memory training research paper

Enhancing Cognitive Abilities with Comprehensive Training: A Large, Online, Randomized, Active-Controlled Trial

Affiliations.

• 1 Department of Research and Development, Lumos Labs, San Francisco, California, United States of America.
• 2 Department of Psychology, Wheaton College, Norton, Massachusetts, United States of America.
• 3 Merrill Palmer Skillman Institute for Child and Family Development, Wayne State University, Detroit, Michigan, United States of America; Department of Pediatrics, Wayne State University School of Medicine, Detroit, Michigan, United States of America.
• PMID: 26333022
• PMCID: PMC4557999
• DOI: 10.1371/journal.pone.0134467

Background: A variety of studies have demonstrated gains in cognitive ability following cognitive training interventions. However, other studies have not shown such gains, and questions remain regarding the efficacy of specific cognitive training interventions. Cognitive training research often involves programs made up of just one or a few exercises, targeting limited and specific cognitive endpoints. In addition, cognitive training studies typically involve small samples that may be insufficient for reliable measurement of change. Other studies have utilized training periods that were too short to generate reliable gains in cognitive performance.

Methods: The present study evaluated an online cognitive training program comprised of 49 exercises targeting a variety of cognitive capacities. The cognitive training program was compared to an active control condition in which participants completed crossword puzzles. All participants were recruited, trained, and tested online (N = 4,715 fully evaluable participants). Participants in both groups were instructed to complete one approximately 15-minute session at least 5 days per week for 10 weeks.

Results: Participants randomly assigned to the treatment group improved significantly more on the primary outcome measure, an aggregate measure of neuropsychological performance, than did the active control group (Cohen's d effect size = 0.255; 95% confidence interval = [0.198, 0.312]). Treatment participants showed greater improvements than controls on speed of processing, short-term memory, working memory, problem solving, and fluid reasoning assessments. Participants in the treatment group also showed greater improvements on self-reported measures of cognitive functioning, particularly on those items related to concentration compared to the control group (Cohen's d = 0.249; 95% confidence interval = [0.191, 0.306]).

Conclusion: Taken together, these results indicate that a varied training program composed of a number of tasks targeted to different cognitive functions can show transfer to a wide range of untrained measures of cognitive performance.

Trial registration: ClinicalTrials.gov NCT-02367898.

Trial registration: ClinicalTrials.gov NCT02367898 .

Publication types

• Randomized Controlled Trial
• Aged, 80 and over
• Attention / physiology*
• Cognition / physiology*
• Memory, Short-Term / physiology*
• Middle Aged
• Neuropsychological Tests
• Practice, Psychological*
• Problem Solving / physiology*
• Young Adult

Associated data

• ClinicalTrials.gov/NCT02367898

Grants and funding

Advertisement

Can short-term memory be trained?

• Open access
• Published: 27 February 2019
• Volume 47 , pages 1012–1023, ( 2019 )

Cite this article

You have full access to this open access article

• Dennis G. Norris 1 ,
• Jane Hall 1 &
• Susan E. Gathercole 1  

14k Accesses

16 Citations

1 Altmetric

Explore all metrics

Is the capacity of short-term memory fixed, or does it improve with practice? It is already known that training on complex working memory tasks is more likely to transfer to untrained tasks with similar properties, but this approach has not been extended to the more basic short-term memory system responsible for verbal serial recall. Here we investigated this with adaptive training algorithms widely applied in working memory training. Serial recall of visually presented digits was found to improve over the course of 20 training sessions, but this improvement did not extend to recall of either spoken digits or visually presented letters. In contrast, training on a nonserial visual short-term memory color change detection task did transfer to a line orientation change detection task. We suggest that training only generates substantial transfer when the unfamiliar demands of the training activities require the development of novel routines that can then be applied to untrained versions of the same paradigm (Gathercole, Dunning, Holmes, & Norris, 2019 ). In contrast, serial recall of digits is fully supported by the existing verbal short-term memory system and does not require the development of new routines.

Similar content being viewed by others

Spacing and Interleaving Effects Require Distinct Theoretical Bases: a Systematic Review Testing the Cognitive Load and Discriminative-Contrast Hypotheses

Extending cognitive load theory to incorporate working memory resource depletion: evidence from the spacing effect.

No one knows what attention is

Avoid common mistakes on your manuscript.

In this study, we asked whether the capacity of short-term memory (STM) can be improved with practice. In recent years there has been strong interest in the potential of cognitive training programs to enhance mental capacity (Bavelier, Green, Pouget, & Schrater, 2012 ; Simons et al., 2016 ) Many training programs employ complex working memory (WM) activities that combine serial recall of memory sequences with other processing demands. For example, participants may be required to engage in distractor activities interpolated between the presentation of memory items (Chein & Morrison, 2010 ) or to continuously update the sequence of memory items to be remembered (Dahlin, Neely, Larsson, Backman, & Nyberg, 2008 ; Jaeggi, Buschkuehl, Jonides, & Perrig, 2008 ). In these programs, the difficulty of the training task adapts as performance improves with practice. After more than a decade of research in this field, the consensus is that this kind of training generates reliable near transfer to untrained WM tasks with similar task demands. However, there is little far transfer to different activities that are also associated with WM, such as attentional control, reasoning, and learning (Cortese et al., 2015 ; Melby-Lervåg & Hulme, 2013 , 2016 ; Simons et al., 2016 ).

To explain this restricted pattern of transfer we have proposed that training on complex WM tasks involves acquiring a new cognitive skill (Gathercole, Dunning, Holmes, & Norris, 2019 ). The suggestion is that to accomplish the unfamiliar tasks present in most WM training programs, trainees must develop novel routines that coordinate the cognitive processes required. Learning a new routine follows the usual course of the acquisition of any cognitive skill (Anderson, 1982 ; Taatgen, 2013 ). With practice, the routine becomes more efficient and less demanding of general cognitive resources, and performance improves. Transfer will only occur if the routine can be successfully applied to a new task, and this will only happen if the task demands are closely matched. Consistent with this framework, a meta-analysis of WM training studies showed that substantial transfer following WM training is largely restricted to cases in which both the trained and untrained tasks share the same complex WM paradigm (Gathercole et al., 2019 ).

We have made the strong claim that new routines will only be developed if existing mechanisms and processes are not available to support the training activities. If existing mechanisms are available, a new routine is not required, and there will therefore be little transfer. It is proposed that verbal STM measures such as digit span are examples of tasks that do not demand a new routine (Gathercole et al., 2019 ). Highly specialized processes in verbal STM are responsible for the encoding, maintenance, and retrieval of phonological material in its original sequence, with key phenomena being successfully simulated by computational models incorporating separate item- and order-encoding mechanisms (Hurlstone, Hitch, & Baddeley, 2014 ). These processes are frequently engaged in everyday situations outside of the laboratory, including learning new words (Baddeley, Gathercole, & Papagno, 1998 ; Gathercole, 2006 ), following complex verbal instructions (Engle, Carullo, & Collins, 1991 ; Jaroslawska, Gathercole, Logie, & Holmes, 2016 ), and performing mental arithmetic (Adams & Hitch, 1997 ; Geary, Hoard, Byrd-Craven, & DeSoto, 2004 ; McLean & Hitch, 1999 ). On this basis, the cognitive-routine framework predicts that, in contrast to complex WM paradigms, training on verbal STM measures will generate little transfer to other verbal serial-recall tasks.

Surprisingly little research has applied the adaptive computerized algorithms widely employed in WM training studies to more simple verbal STM tasks. Older studies involving small numbers of individuals training on single tasks over extended periods have demonstrated that performance on digit serial recall can improve with practice. In the first study of its kind, two adults received more than 50 sessions of digit span testing spread across a period of four months (Martin & Fernberger, 1929 ). Span increased by about 40% in both cases, and this was accompanied by reports of grouping strategies. Striking evidence that memory span gains across training are driven by mnemonic strategies was provided by a study by Chase and Ericsson ( 1981 ) of S.F., an adult whose digit span increased from seven to 79 items over 2 years of practice. This was achieved by recoding digit sequences into long-distance running times that he was familiar with as a long-distance runner. However, the increase in S.F.’s memory span was entirely restricted to digit sequences. When the memory sequences were composed of letters, his span remained at seven across the full training period.

A second long-distance runner, D.D., was instructed to use the same strategy S.F. had developed, and his span also increased substantially (for a detailed analysis of D.D.’s recall strategies, see Ericsson & Staszewski, 1989 ; Staszewski, 1990 ; Yoon, Ericsson, & Donatelli, 2018 ). Other studies have also demonstrated that digit span expands when participants use other elaborate encoding and retrieval strategies, such as the method of loci, the method of mental imagery, and associations between digit sequences and famous historical dates (Kliegl, Smith, Heckhausen, & Baltes, 1987 , and Susukita, 1933, cited in Kliegl et al., 1987 ). Adopting a rather different approach, Reisberg, Rappaport, and O’Shaughnessy ( 1984 ) trained participants to map digits onto finger movements. In this study, digit span increased by up to 50%.

In each of these cases, the increase in span was accompanied by the use of complex mnemonic strategies that combine existing knowledge with either sequences of multiple digits or well-learned sequences that could be used as cues for retrieval. The gains therefore appear to be a consequence of using long-term memory to support the encoding and retrieval of memory items, in conjunction with fixed-capacity verbal STM (Norris, 2017 ). There is little evidence for fundamental capacity changes in verbal STM with extensive practice.

In the only study of its kind, to our knowledge, Harrison et al. ( 2013 ) employed an adaptive computerized STM training regime that consisted of two simple serial tasks—letter span and a spatial span task involving recall of the spatial locations of cells highlighted successively in a matrix. They compared simple span training with two adaptive training programs employing complex span and visual search tasks. Although performance on untrained STM tasks of word span (a verbal serial-recall task) and arrow span (spatial serial recall) improved following training, equivalent benefits were also found for visual search training. There was therefore no selective enhancement of simple memory span by serial-recall training. Other studies have also failed to detect significant transfer of the Cogmed WM training program to digit span, despite its inclusion of a letter span task included in a small number of training sessions (Brehmer, Westerberg, & Bäckman, 2012 ; Dunning & Holmes, 2014 ; Gray et al., 2012 ; Hardy, Willard, Allen, & Bonner, 2013 ).

In the present study, we tested directly whether the capacity of verbal STM can be enhanced through task-specific training of verbal serial recall. Transfer to untrained STM tasks was compared in three groups, each receiving adaptive training on one of the following STM tasks: digit span, circle span, and color change detection. Circle span involves the serial recall of spatial locations highlighted in a sequence at presentation (Minear et al., 2016 ). Like the dot matrix (Alloway, Gathercole, Kirkwood, & Elliott, 2008 ), Corsi block (Darling, Della Sala, Logie, & Cantagallo, 2006 ), and span-board (Wechsler, 1981 ) tasks, circle span is considered to tap a limited-capacity visuospatial STM system (Logie & Pearson, 1997 ). The color change detection task has been widely used as a measure of the capacity of visual STM (e.g., Awh, Barton, & Vogel, 2007 ). Developed by Luck and Vogel ( 1997 ), it involves participants detecting changes in the colors of individual squares presented simultaneously and briefly in a multi-item visual display. It provides the ideal active-control training condition for the two serial-recall training conditions of digit and circle span, as it does not require the retention of serial order. Performance on this task has already been shown to improve with training. With an adaptive algorithm, Buschkuehl, Jaeggi, Mueller, Shah, and Jonides ( 2017 ) reported that set size increased from 6.3 to 8.8 over ten sessions of 300 trials each. In a related nonadaptive task in which only the target square was presented at test, rather than the entire display, Xu, Adam, Fang, and Vogel ( 2018 ) reported an increase in memory capacity from 2.1 to 3.0 across 60 days of training. Harrison et al. ( 2013 ) reported no changes in change detection following training on two complex span tasks involving serial recall. This suggests that the paradigms tap independent STM and WM systems.

The primary aim of the present study was to investigate what changes occur following digit span training. We set out to track transfer by systematically varying the individual features of the trained tasks in a set of untrained tasks administered before and after training. The untrained verbal serial-recall tasks were spoken digit span (a change in presentation modality from the visual trained task) and visual letter span (a change in verbal category). Substantial experimental evidence has shown that each of these input forms gains ready access to the phonological storage component of verbal STM (Baddeley, Lewis, & Vallar, 1984 ).

The design of this study allowed us to address several hypotheses regarding the transferability of STM training. In each case, we did this by comparing the training program of interest with the most appropriate active-control program to test each hypothesis. The first hypothesis was that there would be no transfer following digit span training to other verbal serial-recall tasks, because such tasks depend on specialized processes that are already in place in verbal STM (Gathercole et al., 2019 ). As a consequence, they would not require the development of the novel routines proposed to be the primary source of transfer to untrained tasks with compatible structures. There should therefore be no transfer across verbal serial-recall tasks, even though they place comparable demands on the retention of phonological serial-order information. This should be evident when comparing the impact of digit span training with that of circle span training, the most similar active-control condition, which differed in the domain of STM, but not in the requirement for serial recall. Such an outcome would indicate not only that the fundamental capacity of verbal STM is impervious to training, but also that any substantial on-task training gains are mediated by processes that operate largely outside of STM, such as recoding or chunking (Ericsson, Chase, & Faloon, 1980 ; Martin & Fernberger, 1929 ). Later in this article, we speculate that more subtle improvements may nonetheless result from optimization or fine-tuning of the task model to its unique combination of features.

An alternative hypothesis is that verbal STM can be trained. This would be consistent with claims that the cognitive and neural processes that underpin the broader WM system in which STM is embedded can be modified by intensive training (Astle, Barnes, Baker, Colclough, & Woolrich, 2015 ; Klingberg, 2010 ). If transfer does extend to untrained serial-recall tasks, the key question is: to which tasks? If the serial-order mechanism is both trainable and specific to verbal STM, transfer should not extend beyond verbal serial-recall tasks to any other untrained tasks, including the circle span measure of spatial STM. In this way, the data have the potential to inform long-standing debate about the extent to which the STM mechanisms for retaining serial order are domain-specific or domain-general (Abrahamse, Van Dijck, Majerus, & Fias, 2014 ; Alloway, Gathercole, & Pickering, 2006 ; Bayliss, Jarrold, Gunn, & Baddeley, 2003 ; Engle et al., 1991 ; Hanley, Young, & Pearson, 1991 ; Hurlstone et al., 2014 ; Majerus et al., 2010 ). We asked whether transfer is restricted to serial-recall tasks in the same domain (from digit span to letter span and spoken digit span) or extends across domains (from digit span to circle span or circle span to digit span). Here, the critical comparison to evaluate the specificity of transfer would be digit span training versus nonserial color change training.

To further test the limits on transfer, other untrained tasks were included that did not require memory for serial order. Pattern span involves recall of the pattern of filled cells in a static grid. This measure of visual STM involves the recall of filled cells in a grid that are displayed simultaneously and can be recalled in any order. Under some conditions, performance on this test has been found to be dissociable from serial spatial STM tests such as Corsi block recall, possibly reflecting fractionation of visuospatial STM into separate visual and spatial components (Della Sala, Gray, Baddeley, Allamano, & Wilson, 1999 ). We might therefore anticipate no transfer from either digit span or circle span training (relative to color change detection training) to pattern span.

An untrained line orientation change detection test was also included, in which the array was composed of multiple lines at different orientations and in which participants judged whether the orientation of a single element had changed or remained the same. Alongside color change detection training, this allowed us to test whether training in verbal or spatial serial recall generates benefits that extend to nonserial visual STM. We predicted that it would not, as serial recall appears to reflect a distinct and purely visual system of temporary storage. The inclusion of this training condition also provided an opportunity to explore whether training in one visual change detection task (color change detection) transfers to the ability to detect changes in other visual features in a similar task environment. To date, there has been little research on transfer following training on this paradigm; only a single study has been performed, and this showed no transfer across to variants of the same paradigm with minor changes (Gaspar, Neider, Simons, McCarley, & Kramer, 2013 ). A finding of positive transfer to orientation change detection following color change detection training would provide preliminary evidence for training-related improvement in the ability to detect mismatches in the properties of visual displays, rather than more specifically to detect changes in the colors of individual objects. In the absence of strong hypotheses regarding the outcome, comparisons were made between the color change training group and each of the other three training groups, including a passive control group that received no training, as well as the digit and circle span training groups, which formed the two active controls.

Participants

Eighty English-speaking adults between 18 and 35 years of age were recruited from the Medical Research Council Cognition and Brain Sciences Unit (CBU) Volunteer Panel. Participants received a payment for their time and travel expenses. The study was approved by the University of Cambridge Psychology Research Ethics Committee (CPREC 2014.76). The Participant Information Sheet provided in advance of recruitment outlined a standard hourly payment for participation in the study and reimbursement for travel expenses. It explained that participants would be randomly allocated to conditions, with 10 h of paid home-based iPad training for some but not all individuals. The participants were allocated to training conditions (digit span, circle span, change detection, and no training) on a random basis, subject to the constraint that there were 20 participants in each training condition. This sample size yields power of .91 to detect a large effect size, f 2 = .35 with a general linear regression model (GLM), and .55 to detect a medium effect size, f 2 = .35. The sample size was determined on the basis of the outcomes of a meta-analysis of near transfer following WM training reported by Gathercole et al. ( 2019 ). On the basis of the limited available evidence, they concluded that, in contrast to the robust transfer found following training on complex WM tasks, gains in verbal STM are at best “small in magnitude and may be reliably detected only under conditions of higher statistical power than the standard WM training study” (p. 20). The present sample size of 20 participants per training group falls within the standard range for WM training studies (Dunning & Holmes, 2014 ; Harrison et al., 2013 ; Henry, Messer, & Nash, 2014 ).

Individuals made two 3-h visits to the CBU. On the first visit, each person was given an iPad with a retina display, for use only in the experiment, that provided access to the training program. Participants first completed eight iPad transfer tasks and then eight tests from the Automated Working Memory Assessment (AWMA). The data from the AWMA are reported for completeness only. The training program was then demonstrated, and participants took the iPad away with them to perform the training. The second visit took place shortly after the final session, approximately three weeks later. Participants again completed the iPad and AWMA transfer tests, and then they returned the iPad. All phases of the experiment apart from administration of the AWMA were presented on iPads with a display resolution of 2,048 × 1,536 pixels in landscape mode. At both the pre- and posttraining sessions, there was also a resting-state magnetoencephalography session, the data from which will be reported elsewhere.

Each of the following tasks was administered at the two visits to the CBU, before and after completing the training program. The eight tasks were presented on the iPad, with common designs and structures where possible. After each session of transfer and training, the data were automatically uploaded to a server at the CBU. Two participants did not complete the pretraining line orientation change detection task, and a further participant did not complete the posttraining line orientation task.

Visual digit span

Digits were presented at a rate of one per 750 ms, with each digit being displayed for 500 ms and a blank interval of 250 ms between digits. At the end of the digit sequence, a numeric keyboard (the digits 1–9 in 3 × 3 telephone keypad layout) was displayed, and participants pressed the keys in the order in which the digits had appeared. Below the keyboard was a “Done” key that participants pressed after recall had been completed. With list lengths of nine or less, the digits were sampled randomly without replacement from the digits 1–9. With list lengths greater than nine, the initial set of nine digits was supplemented with a further, randomly sampled N digits. No digits appeared twice in succession, and there were no runs of three or more consecutive ascending or descending digits. Testing began with a block of six trials with a list length of four, and increased by one when participants got four or more of the six trials at that length completely correct. Testing continued until participants failed to reach this continuation threshold. Span was determined to be the longest list length for which four or more lists were recalled correctly. Once span had been determined, 12 further trials were presented at each of the lengths span + 1 and span + 2. The measure of performance was 2 * the number of items recalled correctly during span setting + the numbers of items recalled correctly at span + 1 and span + 2. This gave equal weight to trials at each list length.

Spoken digit span

This task was identical to the visual digit task, except that the stimuli were digits spoken by a male speaker. The digit sound files were padded out with silence to be 500 ms long. Each digit sound file was followed by 250 ms of silence. The method for determining span and the numbers of trials at each of span + 1 and span + 2 were identical to the methods employed for visual digit span.

Letter span

This task was also identical to the visual digit span task, except that the stimulus set was composed of the consonants B, F, H, J, L, M, Q, R, and S. The letters on a 3 × 3 keyboard were arranged in alphabetical order. The method for determining span and the numbers of trials at each of span + 1 and span + 2 were identical to the methods employed for visual digit span.

Circle span

Participants were presented with an array of pseudorandomly positioned circles with a radius of 81 pixels and a minimum center-to-center separation of 272 pixels. All circles were colored medium blue on a gray background and then, in random sequence, each circle turned light blue for 250 ms. The rate of presentation was 750 ms per circle. At the end of the sequence, all circles remained visible, and participants were instructed to touch the circles in the order in which they had been displayed. The method for determining span and the numbers of trials at each of span + 1 and span + 2 were identical to the methods employed for visual digit span. Although this task has been used elsewhere as a test of spatial STM (Minear et al., 2016 ), the possibility cannot be ruled out that participants might attempt to encode the location of the circles verbally. This could introduce an element of verbal STM training in the circle span task. This possibility should be minimized here by the fact that the configuration of the circles varied randomly from trial to trial, making it difficult to assign a set of consistent verbal labels to the locations of the circles.

Pattern span

On each trial a 6 × 6 grid was presented for 500 ms. Initially, four of the cells in the grid were displayed in red. After a delay of 1,000 ms, an empty grid was presented, and participants had to touch the squares that had been presented in red. The squares could be touched in any order. The method for determining span and the numbers of trials at each of span + 1 and span + 2 were identical to the methods employed for visual digit span.

Color change detection

The procedure was based on that used by Luck and Vogel ( 1997 ). Participants saw an array containing between three and 23 colored squares presented for 250 ms. The colors of the squares were chosen at random with replacement from a set of seven readily discriminable colors. The locations of the squares (38 pixels) were random, subject to the constraint that a minimum distance of 117 pixels should separate the centers of the squares.

After a blank retention interval of 1,000 ms, a probe display appeared for 500 ms. The probe was constructed by repeating the previous array, but with one square chosen at random to be the probe square. The color of the probe square either remained the same as in the initial display or, in 50% of the trials, was changed to another randomly chosen color. The location of the probe square was indicated by a red rectangle. Participants had a maximum of 5,000 ms to judge whether the color of the probe square had changed. In all, 20 trials were presented for each of the array sizes 6, 9, and 12. Cowan’s K was used to provide a measure of STM capacity for this task (Cowan, 2001 ; Cowan et al., 2005 ), where K = display size × (proportion hits – proportion false alarms). The mean K was computed over the three array sizes, and this measure was used for the purposes of analysis.

Direction change detection

The orientation change detection task was identical to the color change detection task, except that the colored squares were replaced by black lines that could appear in one of four orientations (vertical, horizontal, or either diagonal). On each probe trial, one line was cued by a red circle, with a 50% probability that the orientation of the line would have changed. The mean K was computed, as in the color change detection task. Due to technical problems, there were incomplete data on this transfer task for two participants in the circle-training condition. Their data were omitted from the reported analyses.

Automated Working Memory Assessment

The following tests from the AWMA (Alloway, 2007 ), a standardized test battery of STM and WM tests, were administered on a desktop PC. Each employed a span procedure. The tests were word span and nonword span tests (verbal STM), dot matrix and mazes memory (visuospatial STM), listening span and counting span (verbal WM), and Mr X and spatial span (visuospatial WM). The analysis was based on raw scores. It should be noted that these tasks shared fewer presentational and task features in common with the most closely matched training activities than did the iPad transfer tasks. On this basis, no strong predictions could be made regarding the impact and potential specificity of the training conditions on these transfer tests. For completeness, the data and statistical outcomes are reported in the supplemental material .

Matrix reasoning

This test of nonverbal reasoning from the Wechsler Adult Intelligence Scales involves selecting the missing part to complete visuospatial patterns. Raw scores were used for the purpose of analysis.

Each participant completed either digit span, circle span, color change detection training, or no training. The participants in the three active training conditions were asked to complete 20 sessions in total on their iPad program. The maximum time allowed for completion of a session was 40 min, with no more than three sessions per day and an interval of no more than two days between successive training sessions. Training could only be performed between 7 a.m. and 11 p.m.

Digit span training

Each training session consisted of eight blocks of ten trials employing the same procedure as the visual digit span transfer test, with the exception that set size was varied adaptively. Training began with a sequence of three digits, and increased by one when participants got eight or more trials completely correct in a block. The length of the sequence decreased by one if participants got two or fewer trials correct. Due to technical problems, the data for one participant in the digit training condition were lost for the 12th session. The missing score was replaced by the mean score from the 11th and 13th sessions. On average, participants completed the training sessions in in 13.8 days (min 10, max 16, SD 1.8). The principal score for the purposes of analysis was the span reached in the final block of each session.

Circle span training

Each training session consisted of eight blocks of ten trials employing the same procedure as the circle span transfer task. Training began with a display of three circles and increased by one when participants got eight or more trials completely correct in a block. The number of circles decreased by one if participants got two or fewer trials correct. On average, participants completed the training sessions in 13.25 days (min 10, max 15, SD 1.55). The principal score for the purposes of analysis was the span reached in the final block of each session.

Color change detection training

In each training session there were eight blocks of 30 trials, employing the same presentation procedure as in the color change detection transfer task. The size of the array was increased by one if participants got 27 or more trials correct, and decreased by one if they got 18 or fewer correct. On average, participants completed the training sessions in 13 days (min 10, max 15, SD 1). Two measures derived from each session were used for the purposes of analysis: The first was the capacity measure K (Cowan, 2001 ; Cowan et al., 2005 ), and the second was the difficulty level (set size) reached by the end of each block.

Analysis plan

The training and transfer data were analyzed using both traditional null-hypothesis significance testing (NHST) methods and corresponding Bayesian methods. This allowed us to quantify the strength of evidence both in favor of the null hypotheses of the absence of training/transfer effects, and in favor of the alternative hypothesis that there were positive effects. The Bayesian analyses were conducted using JASP (JASP Team, 2015 ). Bayes factors (BF 10 ) were interpreted as follows (Jeffreys, 1961 ): BFs < 0.33 provide evidence for the null hypothesis; BFs 0.33–3 provide equivocal evidence for both hypotheses; BFs > 3 provide evidence favoring the alternative hypothesis; BFs > 10 and < 0.01 are considered strong evidence in either direction; and BFs > 100 and < 0.001 provide decisive evidence in either direction.

Training effects for the three active training conditions were analyzed in one-way analyses of variance (ANOVAs) with session as the independent variable. Interactions between training conditions and trials were not computed, due to the different performance metrics used for the span and color change detection tasks. For all training conditions, the metric was the difficulty level achieved by each participant in the final block of each session. For digit and circle span, this was the number of items in the sequence, and for color change detection, it was K . Both Bayesian and non-Bayesian ANOVAs were performed.

To evaluate the specificity of transfer following training, Bayesian and non-Bayesian linear regression analyses were performed for each combination of the training task of interest and each transfer test. The posttraining transfer measure was the dependent variable, and the pretraining measure and the particular training group contrast were entered as dependent variables in each case. Four group contrasts were made for each transfer measure. Three comparisons contrasted pairs of active adaptive training conditions: digit span versus circle span (testing the domain specificity of serial-recall training), digit span versus color change detection (testing the specificity of serial-recall training), and circle span versus color change detection (testing the specificity of change detection training). A final contrast compared color change detection training with no training, as a test of whether there was transfer across serial and nonserial STM paradigms. For the NHST, a Bonferroni correction was applied on a family-wise basis for the transfer tests, yielding an α of .007. For each group contrast and transfer measure combination, initial linear regression analyses were run testing for interactions between pretraining scores and group. Where these were considered to be significant or to favor the alternative hypothesis ( p < .05 or BF > 3), the group term reported here is taken from the analysis that included the interaction term. If the interaction terms did not meet these criteria, the model was rerun excluding the interaction term, and the group term from this analysis is reported.

Training data

The mean scores achieved at the end of each training session (span for digit and circle training, and both capacity K and difficulty level for color change detection) are shown in Fig. 1 . Gains across training sessions were considerably greater for color change detection than for either digit span or circle span training. Performance increased from the first to the final training session, by 18% for digit span training and by 13% for circle span training. For color change detection training, the increase was 51% for the capacity measure K , and 83% for the difficulty level. This reflects an increase in the number of elements in the array from 8.45 to 15.50.

Scores on the final block of each session as a function of training condition and measure

Bayesian and non-Bayesian one-way ANOVAs were performed on the scores in each session for each training condition. In each case, performance increased significantly across training. For NHST, these results were: digit span, F (19, 361) = 7.691, MSE = .497, p < .001; circle span, F (19, 361) = 6.275, MSE = .358, p < .001; color change detection K , F (19, 361) = 8.774, MSE = 1.227, p < .001; color change detection difficulty, F (19, 361) = 50.574, MSE = 1.447, p < .001. Tested against the null model, the BF 10 values were > 100 for digit span, circle span, color change detection K, and color change detection difficulty level. These outcomes provide decisive evidence that performance improved with training on each task and for each measure.

Transfer data

Descriptive statistics and analyses of the transfer measures are shown in Table 1 . Our first question was whether there is domain-specific transfer to other verbal span tasks following digit span training. This was addressed by comparing posttraining scores on the untrained verbal span tests for the digit and circle span training groups. Unsurprisingly, digit span led to a substantially greater enhancement of digit span performance than did circle span training, according to both NHST and Bayesian analysis. This represents a further demonstration of on-task training. For the two untrained verbal span measures, the evidence for a selective advantage following digit span training was weak. For spoken digit span (a change in the modality of the memory items), the NHST was nonsignificant, and the Bayesian outcome was equivocal, weakly favoring the null hypothesis. For letter span, too, the p value was nonsignificant and the Bayes factor value equivocal.

The second question was whether there are domain-general benefits to serial-recall training. This was addressed in two ways. The first was by comparing the digit span and color change detection training groups on the circle span transfer measure. There was no substantial evidence of transfer: The p value was nonsignificant, and the Bayes factor was equivocal, weakly favoring the null hypothesis. The second comparison was between circle span and color change detection training for the three verbal span measures. In each case, the NHST was nonsignificant and the Bayes factor value substantially favored the null hypothesis. We therefore found no substantial evidence for cross-domain transfer across serial-recall tasks in either direction.

The third question was whether there were paradigm-general benefits to training that extended across all three STM training and transfer tasks. This was addressed by comparing the posttraining performance of the color change detection and no-training groups on the five transfer tests that did not involve change detection. It should be noted that this comparison between an active-training and a no-contact control condition was likely to overestimate any potential benefits, and therefore increase the likelihood of a false-positive result (Simons et al., 2016 ). For two of the three verbal span measures (spoken digit span and letter span), the analyses provided no evidence of training benefits, with nonsignificant p values and Bayes factor values showing substantial support for the null hypothesis. For visual digit span, the p value was nonsignificant, and the Bayes factor was equivocal and mildly favored the null hypothesis. For circle span, though, we did find evidence of strong transfer, by both NHST and Bayesian analysis. This provided unexpected evidence for transfer from visual STM training to visuospatial serial recall. However, it is notable that there was no evidence for a symmetrical pattern of transfer from circle span training to the color change detection task: With circle training, capacity K increased from 5.28 to 5.73. For the no-training condition, a similar increase from 5.26 to 5.71 was observed.

The final question was whether color change detection training generates benefits for a line detection task employing the same paradigm. Here the statistical outcomes were clear. Relative to circle span training, color change detection training was associated with greater improvements in the untrained line orientation change detection task, as indicated by a strongly significant p value and a Bayes factor substantially favoring the alternative hypothesis.

Performance improved on all three STM tasks across the course of training. The performance gains across training were relatively small for the two serial-recall tasks—15% for digit span and 12% for circle span training. For color change detection training, the increases were considerably greater, with an estimated STM capacity increase of 51%, and an 83% increase in the size of the array by the end of training.

We observed no positive evidence for transfer from digit span training to circle span, or vice versa. Neither was there strong evidence that digit span training benefited performance on the untrained verbal serial-recall tests of spoken digit span or letter span. The lack of transfer across verbal serial tasks is consistent with the predictions of the cognitive-routine framework (Gathercole et al., 2019 ). According to this theory, transfer occurs only when the demands of the training tasks cannot readily be met by existing STM mechanisms and processes. Only under these conditions will participants need to develop new cognitive routines that control and coordinate the processes involved in performing the task. When training involves only simple verbal serial-recall tasks, no new routines are required because a well-established and highly practiced set of mechanisms is already in place within verbal STM. There should therefore be no substantial transfer, as we indeed found. As was noted by Gathercole et al. ( 2019 ), the exception to this would be individuals with an underdeveloped verbal STM system. In children who do not yet rehearse, rehearsal training does indeed increase memory span (Broadley & MacDonald, 1993 ; Johnston, Johnson, & Gray, 1987 ).

The absence of transfer from either serial-recall training program to untrained serial-recall tasks therefore provides no evidence that the capacity of STM can be expanded with intensive training. This is particularly noteworthy for digit span, as the transfer tasks were distinguished only by a single feature—input modality for spoken digit span (visual to auditory), and semantic category for letter span (digits to letters). According to the current understanding of verbal STM, each of these three stimulus forms should be equally readily represented in verbal STM (Baddeley et al., 1984 ). If training had acted to increase STM capacity, the benefits of this extra capacity should therefore extend to both tasks. One possibility is that training on a single task allows participants to develop category-specific complex recoding strategies that reduce memory load by permit chunking of multi-item sequences (Ericsson et al., 1980 ; Martin & Fernberger, 1929 ). Although this hypothesis may explain the corresponding lack of transfer to letter span, it is not consistent with the corresponding absence of transfer to spoken digit span. With equal access to phonological storage for visual and auditory inputs, any beneficial effects of digit-specific encoding strategies would be expected to extend to digits presented in either modality.

Training-induced changes therefore appear to be tied to the semantic category and input modality of the memory items, as well as to paradigm. Why, then, should performance on the trained task improve at all if, as the absence of transfer suggests, the capacity of verbal STM is unchanged? The present data showed a relatively modest increase of 18% in digit span with training. This is considerably smaller that the gains observed in studies that had explicitly trained digit span strategies involving recoding (Ericsson & Staszewski, 1989 ; Kliegl et al., 1987 ; Reisberg et al., 1984 ; Yoon et al., 2018 ). Moreover, the gains that we found in digit span in the present study appear to be tied to the specific conjunction of the trained task features. One way of explaining this is that extensive training on a single task in which all parameters are fixed (e.g., perceptual, timing, and categorical) allows participants to fine-tune their task model. This could be conceived as a form of learning that takes place within the established system of verbal STM. If performance is finely tuned to all features of a single task, even superficial deviations from the trained task might be sufficient to render the model suboptimal. In this way, subtle changes within an existing system could be detectable in training effects when all task features are preserved, but not generalize to other variants of the same paradigm.

Training on the color change detection measure of visual STM generated both substantial on-task training gains and transfer to another task, in which participants detected changes in the orientation of lines in a multi-item array. To our knowledge, this is the first time that training-induced change has been demonstrated for static visual STM, a resource-limited memory system that has been extensively investigated in recent years (e.g., Alvarez & Cavanagh, 2004 ; Bays, Catalao, & Husain, 2009 ). There are several possible explanations for this outcome. Applying the rationale extended to findings of near transfer following WM training (Dahlin et al., 2008 ; Jaeggi et al., 2008 ; Klingberg, 2010 ), it could be interpreted as reflecting genuine plasticity in the capacity of visual STM. An alternative possibility is that the change detection paradigm may require the establishment of a new cognitive routine (Gathercole et al., 2019 ), and that this is the source of transfer. The very brief presentation of displays containing highly similar objects for a binary change detection judgment certainly imposes highly unfamiliar cognitive demands that quite plausibly might not be met solely through the processes in place within visual STM. The relatively large magnitude of the training gains seen in this task is certainly consistent with mediated learning. Perhaps, then, trainees develop a change detection routine to optimize their performance that—unlike the highly specific tuning to the specific task features seen in digit span training, which shows no substantial generalization—can be readily adapted to the untrained orientation detection task with its very similar demands.

Alternatively, transfer across change detection tasks reflects learning by the participant about the statistical properties of the displays. Such learning might underpin the robust training and transfer gains found for change detection tasks. Orhan and Jacobs ( 2014 ) have suggested that the apparent capacity limitations in visual STM might be due to a mismatch between the participant’s internal model and the true statistics of the stimuli. For example, our change detection tasks had a statistical structure of the elements within the display: Colors were not positioned at random but were constrained to have a minimal separation. This became even more constraining as the number of stimuli in the display increased. When the maximum number of items were in the display, they were closely packed, and there was much less room for variation in position than with fewer items. The orientation change task had the same statistical properties. Perhaps, then, learning about the statistics of the displays in one task could readily transfer to the other. An important question as yet unanswered is whether transfer to other change detection tasks would persist if the statistics of the displays changed between the trained and untrained activities.

Participants might also learn about the characteristics of their internal representations of stimuli in change detection tasks. To optimize the readout of information from memory, participants need to have an accurate model of the internal representation that will be produced by a particular input. In Bayesian terms, they need to develop an accurate generative model of the task. This form of learning or adaptation is likely to be tied to the low-level perceptual properties of the stimuli. In the case of serial recall with letters or digits, this might involve nothing more than fine-tuning, but for a completely novel task like change detection, more work might need to be done.

In summary, on-task performance improves after extensive practice with serial recall of visually presented digits. However, there is little evidence that this improvement confers any advantage to recall of visually presented letters, auditorily presented digits, or sequences of spatial locations. Changing either the stimulus domain, the presentation modality, or the category of the memory items eliminated the benefits of training. Digit span training does not substantially improve the capacity of verbal STM. In contrast, training on an unfamiliar color change detection task produces large gains in performance that transfer to a line orientation change detection task. The large improvement in change detection was unexpected, as change detection is often used to estimate core visual STM capacity. This might have been a consequence of learning how to perform a novel task, in much the same way as for more complex WM tasks, and also how to optimize performance by exploiting the statistical properties of the displays.

Abrahamse, E., Van Dijck, J.-P., Majerus, S., & Fias, W. (2014). Finding the answer in space: The mental whiteboard hypothesis on serial order in working memory. Frontiers in Human Neuroscience , 8 , 932. https://doi.org/10.3389/fnhum.2014.00932

Adams, J. W., & Hitch, G. J. (1997). Working memory and children’s mental addition. Journal of Experimental Child Psychology , 67 , 21–38. https://doi.org/10.1006/jecp.1997.2397

Article   PubMed   Google Scholar  

Alloway, T. P. (2007). Automated working memory assessment. San Antonio: Pearson.

Google Scholar  

Alloway, T. P., Gathercole, S. E., Kirkwood, H., & Elliott, J. (2008). Evaluating the validity of the automated working memory assessment. Educational Psychology , 28 , 725–734.

Article   Google Scholar  

Alloway, T. P., Gathercole, S. E., & Pickering, S. J. (2006). Verbal and visuospatial short-term and working memory in children: are they separable? Child Development , 77 , 1698–1716. https://doi.org/10.1111/j.1467-8624.2006.00968.x

Alvarez, G. A., & Cavanagh, P. (2004). The capacity of visual short-term memory is set both by visual information load and by number of objects. Psychological Science , 15 , 106–111. https://doi.org/10.1111/j.0963-7214.2004.01502006.x

Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological Review , 89 , 369–406. https://doi.org/10.1037/0033-295X.89.4.369

Astle, D. E., Barnes, J. J., Baker, K., Colclough, G. L., & Woolrich, M. W. (2015). Cognitive training enhances intrinsic brain connectivity in childhood. Journal of Neuroscience , 35 , 6277–6283. https://doi.org/10.1523/JNEUROSCI.4517-14.2015

Awh, E., Barton, B., & Vogel, E. K. (2007). Visual working memory represents a fixed number of items regardless of complexity. Psychological Science , 18 , 622–628.

Baddeley, A. D., Gathercole, S. E., & Papagno, C. (1998). The phonological loop as a language learning device. Psychological Review , 105 , 158–173. https://doi.org/10.1037/0033-295X.105.1.158

Baddeley, A. D., Lewis, V., & Vallar, G. (1984). Exploring the articulatory loop. Quarterly Journal of Experimental Psychology , 36A , 233–252.

Bavelier, D., Green, C. S., Pouget, A., & Schrater, P. (2012). Brain plasticity through the life span: Learning to learn and action video games. Annual Review of Neuroscience , 35 , 391–416. https://doi.org/10.1146/annurev-neuro-060909-152832

Bayliss, D. M., Jarrold, C., Gunn, D. M., & Baddeley, A. D. (2003). The complexities of complex span: Explaining individual differences in working memory in children and adults. Journal of Experimental Psychology: General , 132 , 71–92. https://doi.org/10.1037/0096-3445.132.1.71

Bays, P. M., Catalao, R. F. G., & Husain, M. (2009). The precision of visual working memory is set by allocation of a shared resource. Journal of Vision , 9 (10), 7. https://doi.org/10.1167/9.10.7

Brehmer, Y., Westerberg, H., & Bäckman, L. (2012). Working-memory training in younger and older adults: Training gains, transfer, and maintenance. Frontiers in Human Neuroscience , 6 , 63. https://doi.org/10.3389/fnhum.2012.00063

Article   PubMed   PubMed Central   Google Scholar  

Broadley, I., & MacDonald, J. (1993). Teaching short term memory skills to children with Down syndrome. Down Syndrome Research and Practice , 1 , 56–62.

Buschkuehl, M., Jaeggi, S. M., Mueller, S. T., Shah, P., & Jonides, J. (2017). Training change detection leads to substantial task-specific improvement. Journal of Cognitive Enhancement , 1 , 419–433.

Chase, W. G., & Ericsson, K. A. (1981). Skilled memory. In J. R. Anderson (Ed.), Cognitive skills and their acquisition (pp. 141–189). Hillsdale: Erlbaum.

Chein, J. M., & Morrison, A. B. (2010). Expanding the mind’s workspace: training and transfer effects with a complex working memory span task. Psychonomic Bulletin & Review , 17 , 193–199. https://doi.org/10.3758/PBR.17.2.193

Cortese, S., Ferrin, M., Brandeis, D., Buitelaar, J., Daley, D., Dittmann, R. W., . . . Sonuga-Barke, E. J. S. (2015). Cognitive training for attention-deficit/hyperactivity disorder: Meta-analysis of clinical and neuropsychological outcomes from randomized controlled trials. Journal of the American Academy of Child and Adolescent Psychiatry , 54 , 164–174. https://doi.org/10.1016/j.jaac.2014.12.010

Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences , 24 , 87–114, disc. 114–185. https://doi.org/10.1017/S0140525X01003922

Cowan, N., Elliott, E. M., Saults, J. S., Morey, C. C., Mattox, S., Hismjatullina, A., & Conway, A. R. A. (2005). On the capacity of attention: Its estimation and its role in working memory and cognitive aptitudes. Cognitive Psychology , 51 , 42–100. https://doi.org/10.1016/j.cogpsych.2004.12.001

Dahlin, E., Neely, A. S., Larsson, A., Backman, L., & Nyberg, L. (2008). Transfer of learning after updating training mediated by the striatum. Science , 320 , 1510–1512. https://doi.org/10.1126/science.1155466

Darling, S., Della Sala, S., Logie, R. H., & Cantagallo, A. (2006). Neuropsychological evidence for separating components of visuo-spatial working memory. Journal of Neurology , 253 , 176–180.

Della Sala, S., Gray, C., Baddeley, A., Allamano, N., & Wilson, L. (1999). Pattern span: A tool for unwelding visuo-spatial memory. Neuropsychologia , 37 , 1189–1199.

Dunning, D. L., & Holmes, J. (2014). Does working memory training promote the use of strategies on untrained working memory tasks? Memory & Cognition , 42 , 854–862. https://doi.org/10.3758/s13421-014-0410-5

Engle, R. W., Carullo, J. J., & Collins, K. W. (1991). Individual differences in working memory for comprehension and following directions. Journal of Educational Research , 84 , 253–262.

Ericsson, K., Chase, W. G., & Faloon, S. (1980). Acquisition of a memory skill. Science , 208 , 1181–1182.

Ericsson, K. A., & Staszewski, J. J. (1989). Skilled memory and expertise: Mechanisms of exceptional performance. In D. Klahr & K. Kotovsky (Eds.), Complex information processing: The impact of Herbert A. Simon (Vol. 2, pp. 235–267). Hillsdale: Erlbaum.

Gaspar, J. G., Neider, M. B., Simons, D. J., McCarley, J. S., & Kramer, A. F. (2013). Change detection: Training and transfer. PLoS ONE , 8 , e67781. https://doi.org/10.1371/journal.pone.0067781

Gathercole, S. E. (2006). Nonword repetition and word learning: The nature of the relationship. Applied Psycholinguistics , 27 , 513–543.

Gathercole, S. E., Dunning, D. L., Holmes, J., & Norris, D. (2019). Working memory training involves learning new skills. Journal of Memory and Language , 105 , 19–42. https://doi.org/10.1016/j.jml.2018.10.003

Geary, D. C., Hoard, M. K., Byrd-Craven, J., & DeSoto, M. C. (2004). Strategy choices in simple and complex addition: contributions of working memory and counting knowledge for children with mathematical disability. Journal of Experimental Child Psychology , 88 , 121–151.

Gray, S., Chaban, P., Martinussen, R., Goldberg, R., Gotlieb, H., Kronitz, R., . . . Tannock, R. (2012). Effects of a computerized working memory training program on working memory, attention, and academics in adolescents with severe LD and comorbid ADHD: A randomized controlled trial. Journal of Child Psychology and Psychiatry , 53 , 1277–1284.

Hanley, J. R., Young, A. W., & Pearson, N. A. (1991). Impairment of the visuo-spatial sketch pad. Quarterly Journal of Experimental Psychology , 43A , 101–125. https://doi.org/10.1080/14640749108401001

Hardy, K. K., Willard, V. W., Allen, T. M., & Bonner, M. J. (2013). Working memory training in survivors of pediatric cancer: a randomized pilot study. Psycho-Oncology , 22 , 1856–1865. https://doi.org/10.1002/pon.3222

Harrison, T. L., Shipstead, Z., Hicks, K. L., Hambrick, D. Z., Redick, T. S., & Engle, R. W. (2013). Working memory training may increase working memory capacity but not fluid intelligence. Psychological Science , 24 , 2409–2419. https://doi.org/10.1177/0956797613492984

Henry, L. A., Messer, D. J., & Nash, G. (2014). Testing for near and far transfer effects with a short, face-to-face adaptive working memory training intervention in typical children. Infant and Child Development , 23 , 84–103.

Hurlstone, M. J., Hitch, G. J., & Baddeley, A. D. (2014). Memory for serial order across domains: An overview of the literature and directions for future research. Psychological Bulletin , 140 , 339–373. https://doi.org/10.1037/a0034221

Jaeggi, S. M., Buschkuehl, M., Jonides, J., & Perrig, W. J. (2008). Improving fluid intelligence with training on working memory. Proceedings of the National Academy of Sciences , 105 , 6829–6833. https://doi.org/10.1073/pnas.0801268105

Jaroslawska, A. J., Gathercole, S. E., Logie, M. R., & Holmes, J. (2016). Following instructions in a virtual school: Does working memory play a role? Memory & Cognition , 44 , 580–589.

JASP Team. (2015). JASP (Version 0.6.6) [Computer software]. Retrieved from https://jasp-stats.org/download/

Jeffreys, H. (1961). The theory of probability (3rd). Oxford: Oxford University Press, Clarendon Press.

Johnston, R. S., Johnson, C., & Gray, C. (1987). The emergence of the word length effect in young children: The effects of overt and covert rehearsal. British Journal of Developmental Psychology , 5 , 243–248.

Kliegl, R., Smith, J., Heckhausen, J., & Baltes, P. B. (1987). Mnemonic training for the acquisition of skilled digit memory. Cognition and Instruction , 4 , 203–223.

Klingberg, T. (2010). Training and plasticity of working memory. Trends in Cognitive Sciences , 14 , 317–324. https://doi.org/10.1016/j.tics.2010.05.002

Logie, R., & Pearson, D. G. (1997). The inner eye and the inner scribe of visuo-spatial working memory: Evidence from developmental fractionation. European Journal of Cognitive Psychology , 9 , 241–257.

Luck, S. J., & Vogel, E. K. (1997). The capacity of visual working memory for features and conjunctions. Nature , 390 , 279–281. https://doi.org/10.1038/36846

Majerus, S., D’Argembeau, A., Martinez Perez, T., Belayachi, S., Van der Linden, M., Collette, F., … Maquet, P. (2010). The commonality of neural networks for verbal and visual short-term memory. Journal of Cognitive Neuroscience , 22 , 2570–2593. https://doi.org/10.1162/jocn.2009.21378

Martin, P. R., & Fernberger, S. W. (1929). Improvement in memory span. American Journal of Psychology , 41 , 91–94.

McLean, J. F., & Hitch, G. J. (1999). Working memory impairments in children with specific arithmetic learning difficulties. Journal of Experimental Child Psychology , 74 , 240–260. https://doi.org/10.1006/jecp.1999.2516

Melby-Lervåg, M., & Hulme, C. (2013). Is working memory training effective? A meta-analytic review. Developmental Psychology , 49 , 270–291. https://doi.org/10.1037/a0028228

Melby-Lervåg, M., & Hulme, C. (2016). There is no convincing evidence that working memory training is effective: A reply to Au et al. (2014) and Karbach and Verhaeghen (2014). Psychonomic Bulletin & Review , 23 , 324–330.

Minear, M., Brasher, F., Guerrero, C. B., Brasher, M., Moore, A., & Sukeena, J. (2016). A simultaneous examination of two forms of working memory training: Evidence for near transfer only. Memory & Cognition , 44 , 1014–1037.

Norris, D. (2017). Short-term memory and long-term memory are still different. Psychological Bulletin , 143 , 992–1009.

Orhan, A. E., & Jacobs, R. A. (2014). Are performance limitations in visual short-term memory tasks due to capacity limitations or model mismatch? arXiv preprint. arXiv:1407.0644

Reisberg, D., Rappaport, I., & O’Shaughnessy, M. (1984). Limits of working memory: The digit digit-span. Journal of Experimental Psychology: Learning, Memory, and Cognition , 10 , 203–221. https://doi.org/10.1037/0278-7393.10.2.203

Simons, D. J., Boot, W. R., Charness, N., Gathercole, S. E., Chabris, C. F., Hambrick, D. Z., & Stine-Morrow, E. A. L. (2016). Do “brain-training” programs work? Psychological Science in the Public Interest , 17 , 103–186. https://doi.org/10.1177/1529100616661983

Staszewski, J. J. (1990). Exceptional memory: The influence of practice and knowledge on the development of elaborative encoding strategies. In Interactions among aptitudes, strategies, and knowledge in cognitive performance (pp. 252–285). New York: Springer.

Taatgen, N. A. (2013). The nature and transfer of cognitive skills. Psychological Review , 120 , 439–471. https://doi.org/10.1037/a0033138

Wechsler, D. (1981). WAIS–R manual. New York: Psychological Corp.

Xu, Z., Adam, K., Fang, X., & Vogel, E. (2018). The reliability and stability of visual working memory capacity. Behavior Research Methods , 50 , 576–588.

Yoon, J. S., Ericsson, K. A., & Donatelli, D. (2018). Effects of 30 years of disuse on exceptional memory performance. Cognitive Science , 42 (Suppl. 3), 884–903. https://doi.org/10.1111/cogs.12562

Download references

Author information

Authors and affiliations.

MRC Cognition and Brain Sciences Unit, University of Cambridge, Cambridge, UK

Dennis G. Norris, Jane Hall & Susan E. Gathercole

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Dennis G. Norris .

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Electronic supplementary material

(DOCX 18 kb)

Rights and permissions

OpenAccess This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Reprints and permissions

About this article

Norris, D.G., Hall, J. & Gathercole, S.E. Can short-term memory be trained?. Mem Cogn 47 , 1012–1023 (2019). https://doi.org/10.3758/s13421-019-00901-z

Download citation

Published : 27 February 2019

Issue Date : 15 July 2019

DOI : https://doi.org/10.3758/s13421-019-00901-z

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Short-term memory
• Memory training
• Find a journal
• Publish with us
• Track your research

• Research article
• Open access
• Published: 25 September 2011

Cognitive and memory training in adults at risk of dementia: A Systematic Review

• Nicola J Gates 1 ,
• Perminder S Sachdev 1 , 2 , 4 ,
• Maria A Fiatarone Singh 5 , 6 &
• Michael Valenzuela 1 , 2 , 3  

BMC Geriatrics volume  11 , Article number:  55 ( 2011 ) Cite this article

41k Accesses

187 Citations

5 Altmetric

Metrics details

Effective non-pharmacological cognitive interventions to prevent Alzheimer's dementia or slow its progression are an urgent international priority. The aim of this review was to evaluate cognitive training trials in individuals with mild cognitive impairment (MCI), and evaluate the efficacy of training in memory strategies or cognitive exercises to determine if cognitive training could benefit individuals at risk of developing dementia.

A systematic review of eligible trials was undertaken, followed by effect size analysis. Cognitive training was differentiated from other cognitive interventions not meeting generally accepted definitions, and included both cognitive exercises and memory strategies.

Ten studies enrolling a total of 305 subjects met criteria for cognitive training in MCI. Only five of the studies were randomized controlled trials. Meta-analysis was not considered appropriate due to the heterogeneity of interventions. Moderate effects on memory outcomes were identified in seven trials. Cognitive exercises (relative effect sizes ranged from .10 to 1.21) may lead to greater benefits than memory strategies (.88 to -1.18) on memory.

Conclusions

Previous conclusions of a lack of efficacy for cognitive training in MCI may have been influenced by not clearly defining the intervention. Our systematic review found that cognitive exercises can produce moderate-to-large beneficial effects on memory-related outcomes. However, the number of high quality RCTs remains low, and so further trials must be a priority. Several suggestions for the better design of cognitive training trials are provided.

Peer Review reports

Development of preventative strategies for Alzheimer's dementia (AD) is an international priority, with prevalence rates projected to increase by over 75% in the next quarter of a century [ 1 ]. One approach to reduce the prevalence of AD is to develop strategies to delay its onset in healthy individuals or those at risk of developing dementia. Prospective cohort studies have found that participation in mentally-stimulating activities is associated with a lower incidence of AD [ 2 ] and even late-life mental activity exhibits a dose-dependent inverse relationship with dementia risk, independent of early life experiences [ 3 ]. Consequently, it is possible that participation in complex mental activities at older age may offer protection from cognitive decline and hence mitigate dementia risk.

Cognitive training provides structured practice of complex mental activity in order to enhance cognitive function [ 4 ], and has attracted intense public, commercial and scientific interest. Unfortunately, cognitive training interventions have been frequently mislabelled or conflated with other therapies, despite important theoretical distinctions between compensatory cognitive rehabilitation, general cognitive stimulation and cognitive training [ 5 – 7 ]. For example, the non-specific umbrella terms 'cognitive intervention' [ 8 ], 'cognitive enrichment' [ 9 ] and 'cognitive rehabilitation' have been applied to multidomain cognitive training [ 10 – 12 ] as well as training in memory strategies [ 13 ]. 'Cognitive stimulation' has been used to refer to interventions ranging from generic topical discussions [ 14 ], executive exercises and memory strategy training [ 15 ]. Given the confusion of terms, an operational definition has been advanced which delineates cognitive training from other interventions [ 16 ]: 1) repeated practice, 2) on problem activities, 3) using standardized tasks, and 4) that target specified cognitive domains.

Cognitive training can be further distinguished to include training in applied memory strategies versus repetitive cognitive exercises [ 7 ]. Training in memory strategies involves the instruction and practice of techniques to minimize memory impairment and enhance performance, and involves learning and practicing strategies such as the method of loci, mnemonics, and visual imagery [ 17 , 18 ]. In contrast, cognitive exercise requires the repeated practice of targeted cognitive abilities in a repetitions-sessions format analogous to 'reps-sets' regimes in physical resistance training: users typically carry out a number of iterations of a cognitive task in one session, then continue to new tasks in the next session, and eventually return to further train the original task at a harder level in future sessions (i.e., staircase design). Recently, several software applications have been developed that implement cognitive exercises on computer [ 19 , 20 ].

Although cognitive exercises and memory strategies are structurally distinct, they have often been analysed together. A Cochrane review of 32 training trials up to the year 2007, concluded that none of the effects could be attributed specifically to cognitive training, however, only memory training data from 24 trials were pooled for analysis, and the analysis did not include results from cognitive exercise trials of problem solving and speed of information processing [ 4 ]. Similarly a review of memory strategy training in healthy and mild cognitive impairment (MCI) individuals [ 18 ] combined results from two trials of cognitive exercises [ 11 , 19 ] with 22 trials of memory strategy training and found no specific effects of training. Furthermore, mixed results were also obtained in a systematic review of cognitive interventions in MCI which included training in both memory strategies and cognitive exercises [ 21 ]. In addition, many of the trials included uncontrolled interventions such as use of external memory aids or relaxation therapy [ 22 ]. Prior reviews have therefore not appropriately distinguished between types of cognitive training, potentially obscuring clinically-relevant effects. Furthermore, a lack of differentiation between cognitive exercises and training in memory strategies, and the inclusion of multiple other therapies with cognitive training, may have also contributed to mixed findings.

By contrast, a meta-analysis of longitudinal RCTs of cognitive training (as defined here) in cognitively healthy adults demonstrated efficacy on primary cognitive outcomes [ 23 ]. However, whether operationally-defined cognitive training can be as effective at slowing the rate of cognitive decline after clinical signs are apparent is not clear. MCI is a diagnostic term applied to those individuals with high risk of developing dementia and in the intermediate stage between normal cognitive function and dementia [ 24 , 25 ]. MCI increases the risk for dementia, with diagnosed individuals progressing at rates of 12-15% per year compared to 1-2% of the general population [ 26 ]. Cognitive training at this preclinical stage may potentially prevent or delay disease onset, reducing this high conversion rate.

The purpose of this systematic review was therefore to identify all relevant clinical trials of defined cognitive training in individuals with MCI in order to: a) determine the overall efficacy of cognitive training in at risk individuals; b) compare outcomes between cognitive exercises and memory strategy training; c) examine the issue of generalisation of training; d) identify and discuss limitations of current research, and e) provide recommendations for future research.

Search Strategy

To identify relevant research trials Medline (1996-18 March 2011), EMBASE (1980-18 March 2011), CINAHL (1980-18 March 2011) and PsychINFO (1984-18 March 2011) databases were searched by NG. The key search term ["cognitive training"] was supplemented with ["cognitive intervention"] ["cognitive rehabilitation"] ["cognitive stimulation"] [cognitive enrichment'] ["memory training"] and ["memory rehabilitation"]. The sample population of interest was the elderly with cognitive impairment but no dementia, and in order to identify this group multiple search terms were entered ["MCI" or "mild cognitive impairment"][ "pre-dementia"] ["mild cognitive disorder"] [" age associated cognitive decline"]or ["cognitive impairment no dementia"]. Combined intervention and population terms were searched in "All Fields", and identified papers were reviewed (title/abstract) by NG to identify potentially relevant trials and this was supplemented by reviewing the reference list of retrieved trials.

Inclusion criteria

Studies were selected from the initial search if they met the following criteria i) described a cognitive training intervention consistent with our definition ii) were a full length article published in a peer reviewed English language journal iii) study design was a randomised controlled trial (RCT), or non-randomised (NRCT) or uncontrolled clinical trial (UCT) iv) sample population was defined as having MCI or in a mixed sample the data for MCI was available separately [ 16 ] v) no training in external memory aids and vi) baseline and post intervention results on at least one cognitive outcome measure.

Appraisal of Study Quality and Data extraction

Included studies were individually scored on their published adherence to the CONSORT 2001 reporting criteria for clinical trials ( http://www.consort-statement.org ) accessed 17 March 2011). Key information was extracted by two reviewers (NG and MV) onto a standard template and any differences resolved by consensus with other authors. Additional non published outcome data were received from the authors of two trials [ 27 , 28 ].

Analytical Approach

Data were extracted for the description of methodology and common outcomes of each trial. A quantitative meta-analysis was our primary goal if a sufficient number of quality studies using homogeneous interventions and outcomes were identified. Our secondary goal was to calculate effect sizes, statistical power along with clinical relevance in all studies and describe important trial characteristics such as: cohort, intervention type, training delivery, volume of training, outcome measures, and follow-up. Relative effect sizes for RCTs and NRCTs were calculated as a difference of change scores with pooled baseline standard deviation (Coe's Calculator retrieved May 5, 2009 from http://www.cemcentre.org/evidence-based-education/effect-size-calculator ). Hedge's bias corrected relative effect sizes were obtained with 95% confidence intervals as this method adjusts for small sample size. Post-hoc power calculations were calculated with GPower Analysis Version 2.0 [ 29 ].

Search Results

A summary of eligible articles into the review is presented in Figure 1 . The combined search of intervention terms AND population terms yielded 175 potentially eligible papers. The abstracts were reviewed, providing a final total of 34 studies, agreed upon by all authors, which were reviewed in full to determine suitability for inclusion. Ten studies met our criteria for cognitive training and MCI: 6 trials of cognitive exercises (3 RCTs, 2 UCTs, 1 NRCT) and 4 training in memory strategies (2 RCTs, 2 NRCTs). A number of studies were excluded because the sample was mixed [ 30 , 31 ] or not defined as MCI [ 32 ], applied individualised non-standardized training [ 33 ] or the intervention combined multiple therapies [ 22 , 34 , 35 ].

Flow of eligible trials into review .

A meta-analysis of the RCTs was considered inappropriate as not more than two studies shared the same global outcome measure (Mini Mental State Exam (MMSE) [ 11 , 27 ]), memory specific story or paragraph recall [ 28 , 36 ], and no common mood outcome measure. Outcome measures were similarly heterogeneous in NRCTs and UCTs. The ten trials in MCI were analysed individually with sample characteristics described in Table 1 , intervention and outcomes described in Table 2 , and effect sizes presented in Tables 3 and 4 .

Study Quality Assessment

There was significant disparity in RCT study quality, with an average CONSORT rating of only 13.5 out of a possible 22 items, with limitations primarily due to poorly explicated methodology and randomization process. The highest scores of 18 and 17 were obtained from a multi-component cognitive exercise intervention [ 27 ] and computer delivered cognitive exercise [ 37 ] respectively, and the remaining 3 trials were awarded scores of between 9 and 11.

Cohort Characteristics

Summary descriptions of each study cohort are shown in Table 1 .

The ten trials yielded a total of 305 participants, with small cohorts ranging in size from 10 [ 8 ] to 59 [ 11 ]. All participants were community-dwelling individuals. Recruitment source was variable, with referral from geriatric, psychiatric, memory clinics or neurology units most common [ 8 , 10 , 15 , 27 , 37 ], however frequently no recruitment information was reported [ 11 , 12 , 36 ]. Disparity in the type and quality of reported demographic information precluded mean calculations. Participants were predominantly women aged in the mid-seventies who had completed secondary school. In some studies, there was an inequality between treatment and control groups at baseline in gender ratios [ 36 ], and MMSE scores [ 27 , 28 ]. Limited information was provided regarding health status and medication, and only five studies listed exclusion criteria [ 8 , 11 , 15 , 27 , 37 ].

MCI diagnosis

Nine trials applied formal diagnostic criteria to determine the MCI status of subjects. Petersen's MCI criteria were most commonly adopted [ 8 , 11 , 12 , 15 , 28 , 36 , 38 ] indicating a predominance of MCI amnesic subtype. No studies operationalized the first criterion of subjective memory complaint [ 39 ], although complaints were assessed during interview [ 28 ]. In contrast, objective operational measures for the remaining 3 criteria of objective memory impairment, intact general cognitive function, and no functional impairments were uniformly provided. All but two studies, [ 37 , 38 ] provided MMSE scores as a measure of baseline cognitive function, with an average score of 26.32 across eight trials, and range of 28.9 [ 8 ] to 17.2 [ 27 ] suggesting significant disparity in level of cognitive impairment.

Cognitive Training Intervention

Training format and delivery.

Training characteristics are presented in Table 2 . Computerized exercises were the most common form of training. The computer programs NeuroPsychological Training (NPT) [ 40 ] and Cogpack [ 41 ] provided multi-modal and multiple-domain training [ 10 – 12 , 38 ], whilst the POSIT Science Corporation (San Francisco, CA) program trained only one cognitive domain, i.e. auditory processing [ 37 ]. One cognitive exercise trial included pen and paper tasks of repeated 30 minute cancellation, ordering, and mathematical tasks [ 27 ].

In contrast memory strategy training involved written and verbal practice of memory strategies including visual imagery [ 8 , 15 , 36 ], association or categorization [ 8 , 15 , 28 , 36 ] and spaced retrieval [ 28 ].

Training delivery for both cognitive exercises and memory strategies most often occurred in group [ 8 , 12 , 15 , 28 , 36 ], or combined group and individual sessions [ 27 ]. One cognitive exercise trial exclusively involved individual home training [ 37 ]. Two memory strategy studies included homework exercises [ 28 , 36 ]. The four memory strategy trials reported that training was supervised by psychologists or neuropsychologists [ 8 , 15 , 28 , 36 ].

Volume and Duration

Volume of cognitive training measured by hours per week was variable, ranging from 1 hour of memory strategy training [ 28 ] to over 8 hours of cognitive exercises [ 37 ]. In trials of combined interventions, it was difficult to delineate duration of training from the other intervention components. Duration of exercise training varied from 3 weeks [ 12 ] to up to 1 year [ 27 ], and memory strategy training ranged from 6 [ 36 ] to 26 weeks [ 28 ]. Overall mean volume of training (sessions/week × number of weeks) was 8 sessions for memory strategies and 57.5 sessions for cognitive exercises.

Combined intervention

Interpretation of results was confounded in all four memory strategy trials [ 8 , 28 , 36 ], as the predominant memory strategy training was augmented with other interventions that were not controlled for in the comparison group. For example, memory strategy training was combined with occupational therapy and behavioural training [ 36 ], life-style education [ 28 ], computer assisted attention training [ 8 ], and executive exercises [ 15 ]. Only one cognitive exercise trial combined other therapy, motor and daily living function training [ 27 ].

Outcome Measures

There was considerable variability in the type and quality of outcome measures used in each study, limiting the extent to which the efficacy of cognitive training in MCI could be evaluated. The types of outcome measures employed can be broadly classified into training-specific measures, domain-specific cognitive measures (memory, attention, executive function, and speed), global cognitive measures, and secondary generalization measures of function and emotional and behavioural status. Word lists and story or paragraph recall were the most common domain-specific outcome measures. There were no measures of quality of life, and incident dementia was not reported in any trial.

Effect Size Analysis

Eight of ten studies reported improvement in at least one cognitive outcome (see Tables 3 and 4 ). Specifically, relative effect sizes (ES) varied between moderate (ES 0.3 - 0.5) to large (> 0.5) on measures of objective memory performance in four of five RCTs (3 cognitive exercise trials [ 11 , 27 , 37 ] and 1 memory strategy trial [ 36 ]), one of three NRCTs [ 8 ], and both uncontrolled cognitive exercise trials [ 10 , 38 ]. However, many results were not statistically significant and so post-hoc power calculations were used to assess the rate of probable Type II errors. All RCTs and NRCTs were found to be underpowered (power less than 80%) for the reported memory outcomes.

Effect sizes on memory outcomes were typically greater for the randomized cognitive exercise trials with relative effect sizes on memory outcomes ranging from .10 to 1.21. In contrast relative effect sizes on memory outcomes following memory strategy training ranged from .88 to -1.18. On tests of text recall, this difference appeared strongest. For example, Rozzini et al 2007 found relative ESs that ranged from ES = .82 to ES = .99 based on story recall [ 11 ], whilst a study of memory strategy training based on paragraph recall failed to find positive training effects (ES = -.03 to ES = -.54 [ 36 ]).

Two RCTs of memory strategy training [ 28 , 36 ] have yielded mixed ES on memory. These studies found no evidence of generalization, with effects being restricted to training-specific outcome measures. Although persistence of strategy use was reported [ 28 ], memory performance deteriorated over time [ 28 , 36 ]. In addition, trained subjects had less improvement at six months than control subjects (ES = -1.18) [ 36 ]. Overall, large effects (ES > .05) were found for 50% of the memory outcomes in cognitive exercise trials compared to 37% of memory strategy outcomes.

Improvements in mood following cognitive exercises were also found. Cognitive exercises led to a reduction in depressive symptoms in both RCT trials [ 11 , 27 ], suggesting that training benefits may include improved mood. Furthermore a large, significant and clinically relevant reduction in psychiatric and depressive symptoms found after cognitive exercises compared to Cholinesterase inhibitors (ChEI) treatment (ES = -.82) [ 11 ], suggesting that cognitive exercise training may have adjunctive benefit to this medication. The two randomized memory strategy trials did not include outcome data on measures of mood [ 28 , 36 ].

Results from the NRCT and UCTs indicated larger effects and clinical benefits following computer-based exercises compared to pen-and-paper memory strategy training across a number of domain specific, global cognitive, and mood function measures [ 10 , 12 , 38 ]. For example, an UCT [ 10 ] and NRCT [ 12 ] cognitive exercise trials both resulted in a small reduction of depressive symptoms. Belleville's 2008 NRCT of memory strategy training yielded a large positive result on word list recall however, interpretation is difficult due to lack of randomization and the use of multiple interventions, including computer exercises, with the relative contribution of each intervention not discernable.

Greater volume of training was associated with greater effect on memory outcomes following cognitive exercises (60 sessions ES = .82 and .99 [ 11 ], compared to 12 sessions ES = .10 [ 12 ]), but not for memory strategy training (6 sessions ES = .88 [ 36 ], compared to 10 sessions ES = .23 [ 28 ]). Combined memory strategy with attention training [ 8 ] was associated with large and significant benefit (ES = .78) following only 8 sessions. Supervision of training was reported in the four memory strategy training trials [ 8 , 15 , 28 , 36 ], making it impossible to evaluate the benefit of supervision compared to no supervision in strategy training, and was not reported upon in the computer-based cognitive exercises trials. Training was provided in group format with the exception of two trials [ 11 , 37 ] therefore comparison of benefit between individual versus group format was not feasible.

Longitudinal Follow-up

Three RCT studies examined persistence of effect with longitudinal follow-up. Unexpectedly, memory strategy training was associated with decreases in objective memory performance at three [ 28 ] and six months [ 36 ] post training compared to control. By contrast, improvements in function were evident following cognitive exercises at three months [ 11 ]. However, an uncontrolled cognitive exercise trial found the benefit of training did not persist at five months follow-up [ 38 ].

This systematic review applied defined criteria to identify original cognitive training studies that investigated cognitive efficacy in preclinical MCI subjects. In addition to RCTs, less robust NRCT and UCT designs were included in this review to exhaustively review the existing literature, the relationship of study quality to outcomes achieved, and identify gaps in the literature. Nonetheless, our systematic review identified only ten trials of cognitive training using either cognitive exercise or memory strategy approaches, of which half were RCTs of low to moderate quality, with significant heterogeneity. Despite these limitations, certain patterns did emerge. Moderate-sized effects were found on memory performance and global cognitive measures in a majority of studies, with computer-based cognitive exercise studies exhibiting an increased frequency of stronger effect sizes, and enhanced generalization of benefits, compared to memory strategy training. However, most studies were underpowered for the effects achieved, and so individual results were often insignificant. Furthermore, three trials [ 27 , 28 , 36 ] included additional intervention components so that the unique benefit of cognitive training is difficult to assess. Overall, the field is nascent and further high quality RCTs are of critical importance.

Defining and Classifying Cognitive Training

The literature regarding cognitive training has so far suffered from a variable definition of intervention, as well as the frequent use of multiple interventions without appropriate controls, thereby accounting for the inconsistent results. For example, two recent meta-analyses of cognitive training in healthy older adults [ 23 , 42 ] drew different conclusions, mainly because inclusion criteria varied, with only three studies being common to both analyses. In MCI, previous reviews have included mixed interventions without clearly delineating between cognitive exercise and memory strategy training, and also included different trials [ 4 , 43 , 44 ]. Mixed interventions only add to the confusion. For example, a recent randomized control trial nominally compared 'cognitive training', comprising health education, meditation, memory strategy training and problem solving strategies, with 'cognitive stimulation', itself comprised of reality orientation, quiz games and problem solving activities [ 45 ]. In the area of cognitive training in established dementia, reviews have similarly produced mixed results: two reports found little support for efficacy [ 46 , 47 ], but a subsequent review which distinguished between restorative cognitive training and compensatory strategies found that cognitive training appeared to be of modest benefit [ 48 ]. If general stimulation activities and rehabilitative compensation are excluded, perhaps, cognitive training can be examined appropriately as a unique stand-alone intervention. In addition, we here distinguished between cognitive exercises and memory strategy training because of fundamental differences in approach and intent.

Evidence from trials of Cognitive training in MCI

Cognitive exercise involving multiple cognitive domains appears to demonstrate greater efficacy than uni-modal memory strategy training. Multi-domain exercises provide a broader range of cognitive challenges to directly stimulate plasticity, and in several studies has resulted in improved global cognitive function [ 11 , 12 , 27 , 49 ]. By contrast, little evidence was found for the efficacy of memory strategy training in MCI which was consistent with outcomes from a recent meta-analysis in healthy and MCI subjects that found training effects were equivalent to those seen in active controls [ 50 ]. Memory strategy training may have limited generalizability to overall cognitive function, perhaps because it has a very specific nature and reliance on subjects' ability to appropriately apply acquired strategies. Since complex mental activity induces a number of central nervous system adaptations, including neuro-protective, plastic, trophic and compensatory mechanisms [ 50 ], multi-domain cognitive exercise may be better suited to stimulating these neuroplastic brain changes. Neuroimaging studies, for example, are beginning to isolate functional, structural and biochemical changes that accompany cognitive training [Suo & Valenzuela, in press]. Further research is required that directly compares memory strategy training to multi-modal cognitive exercises, as well as single- versus multi-domain cognitive exercises. For the moment it is important to recognise that firm conclusions cannot be made because of the prevalence of mixed interventions and overall limited quality studies.

High volume cognitive exercise appeared to result in greater benefit than lower volumes of training, although no dose-response studies were identified. Very frequent training for twelve weeks led to greater effect on memory [ 11 ] than longer, less regular training [ 27 , 38 ], and cognitive exercise studies generally had a higher frequency of training sessions at four [ 10 , 12 ] or five sessions per week [ 11 , 37 ]. The observation of greater efficacy from cognitive exercise than memory strategy training may therefore simply represent a function of training volume, and so volume-matched comparative studies are required. Meta-analysis of cognitive training in healthy adults has suggested that 2-3 month training periods may have persistent protective benefit [ 23 ] however current findings in MCI suggest that frequency and total volume of sessions are also important. Accordingly, it is possible that training needs to reach a 'critical threshold' in order to produce sufficient adaptive neurobiological changes. However, given the variability observed, it is not yet possible to determine the minimum required frequency, volume or duration of cognitive training and explicit dose-response trials in MCI that are needed.

Clinical Role of CT in Primary and Secondary Prevention

Three stages of AD prevention have been identified: primary prevention to reduce disease incidence in cognitively normal individuals; secondary prevention to slow progression of pre-clinical disease to clinical disease (often translating to reduction of MCI 'conversion' to dementia); and tertiary prevention , the reduction of disability due to cognitive symptoms in diagnosed patients [ 51 ]. Cognitive training can be applied to each of these stages, and it is proposed that the type of training intervention should vary depending upon the prevention stage. Cognitive exercise is designed to improve function through neuroplastic mechanisms and has been shown to produce positive effects in healthy adults [ 23 ] thereby consistent with a primary prevention goal. This current review suggests cognitive exercise also has promise for enhancing cognitive function in MCI [ 10 – 12 , 37 , 38 ], and may slow decline in at risk individuals, consistent with secondary prevention. As the syndrome of MCI may have different aetiologies [ 25 ], it is likely that individuals with different subtypes of MCI may respond differently to treatment. For some, training in strategies to compensate for memory difficulties may have additional value, and consequently combined cognitive exercise and memory strategy training may be optimal. However, comparative trials are required and there remains a lack of longitudinal research. The effectiveness of cognitive exercise as a tertiary prevention in those with established AD is likely to be modest [ 48 ], although a recent trial of computer-based exercises found delayed progression of disease by the end of training compared to controls [ 52 ]. For tertiary prevention, compensatory rehabilitative memory strategy training approaches that target disability maybe appropriate.

Computer-delivered interventions are rapidly becoming popular. Computerized cognitive exercise has been successfully implemented across the age spectrum and research suggests that older adults are often the fastest growing users of computer and internet technology [ 53 ]. Computer delivered exercises may provide primary and secondary prevention and be accessible to a wide number of individuals. On the other hand, it may be more appropriate to deliver tertiary compensation strategies by traditional pen and paper methods that minimize memory load. Finally, across all of these types of interventions and prevention stages, quantifying the 'real world' significance of cognitive training interventions on instrumental activities of daily living and quality of life will be vital and so far has not been addressed.

Further research is urgently required in order to substantiate the efficacy of cognitive training as a therapeutic intervention in MCI. It is vital to clearly distinguish between various cognitive interventions and differentiate between training exercises and memory strategies. This review suggests cognitive exercise may be effective at enhancing cognitive outcomes, but several limitations have been identified which precludes firm conclusions. All trials have been small and generally underpowered, and thus larger and more diverse cohorts are needed. Notably, only two RCTs [ 11 , 37 ] to date have examined the isolated benefit of cognitive exercise, whilst the other three included co-interventions. Importantly, no significant negative or adverse effects of cognitive training have been found, in marked contrast to drug trials in MCI, where side effects and high dropout rates are commonly reported [ 47 ].

The following recommendations are intended as an indicative rather than exhaustive list, and demonstrate the challenges and opportunities for the field:

1. Employ a rigorous cognitive training definition, and distinguish between cognitive exercises and memory strategy training in abstracts and reports. A consistent operational definition will facilitate the appropriate comparison of effects between interventions, and provide more precise information for program development and research.

2. RCTs with active control conditions (sham training) are required to control for non-specific effects. Similarly, in multi-modal and combined interventions, additional study arms are needed.

3. Assess generalization by testing cognitive, behavioural, quality of life, functional, mood, and psychological wellbeing outcomes.

4. Compare the effects of training volume and duration, as well as investigate dose-response relationships.

5. Proximal and longitudinal follow-up assessments are needed to determine the persistence of effects and begin characterising the temporal course of putative benefits.

6. Comprehensive description of inclusion criteria and sample descriptors are needed in order to control for the potential heterogeneity in MCI aetiology.

Wimo A, Jonsson L, Winbald B: An estimate of the worldwide prevalence and direct costs of dementia in 2003. Dementia Geriatr Cogn Disorders. 2006, 21: 175-181. 10.1159/000090733.

Article   CAS   Google Scholar  

Wilson R, Mendes de Leon C, Barnes L: Participation in cognitive stimulating activities and risk of incident Alzheimer's disease. JAMA. 2002, 287: 742-748. 10.1001/jama.287.6.742.

Article   PubMed   Google Scholar  

Valenzuela MJ, Sachdev P: Brain Reserve and cognitive decline: a non parametric systematic review. Psychological Medicine. 2006, 36: 1065-1073. 10.1017/S0033291706007744.

Martin M, Clare L, Altgassen AM, Cameron M, Zehnder F: Cognition-based interventions for healthy older people and people with mild cognitive impairment. Cochrane Database of Systematic Reviews. 2011, 1

Belleville S: Cognitive training for persons with mild cognitive impairment. International Psychogeriatrics. 2008, 20 (1): 57-66.

Clare L: Cognitive training and cognitive rehabilitation for people with early-stage dementia. Reviews in Clinical Gerontology. 2003, 13 (1): 75-83.

Article   Google Scholar  

Mowszowski L, Batchelor J, Naismith S: Early intervention for cognitive decline: can cognitive training be used as a selective prevention technique?. International Psychogeriatrics. 2010, 22 (4): 537-548. 10.1017/S1041610209991748.

Belleville S, Gilbert B, Fontaine F, Gagnon L, Menard E, Gauthier S: Improvement of episodic memory in persons with mild cognitive impairment and healthy older adults: evidence from a cognitive intervention program. Dementia & Geriatric Cognitive Disorders. 2006, 22 (5-6): 486-499. 10.1159/000096316.

Schmiedek F, Bauer C, Lovden M, Brose A, Lindenberger U: Cognitive enrichment in old age: web-based trainng programs. Gero Psych. 2010, 23 (2): 59-67.

Google Scholar  

Cipriani G, Bianchette A, Trabucchi M: Outcomes of a computer-based cognitive rehabilitation program on Alzheimer's disease patients compared with those on patients affected by mild cohnitive impairment. Archives of Gerontology and Geriatrics. 2006, 43: 327-335. 10.1016/j.archger.2005.12.003.

Rozzini L, Costardi D, Chilovi V, Franzoni S, Trabucchi M, Padovani A: Efficacy of cognitive rehabilitation in patients with mild cognitive impairment treated with cholinesterase inhibitors. International Journal of Geriatric Psychiatry. 2007, 22 (4): 356-360. 10.1002/gps.1681.

Talassi E, Guerreschi M, Feriani M, Fedi V, Bianchetti A, Trabucchi M: Effectiveness of a cognitive rehabilitation program in mild dementia (MD) and mild cognitive impairment (MCI): A case control study. Archives of Gerontology & Geriatrics. 2007, suppl (1): 391-399.

Loewenstein D, Acevedo A, Czaja S, Duara R: Cognitive Rehabilitation of Mildly Impaired Alzheimer Disease Patients on Cholinesterase Inhibitors. SO - American Journal of Geriatric Psychiatry July/August 2004. 2004, 12 (4): 395-402.

Tarraga L, Boada M, Modinos G: A randomized pilot study to assess the efficacy of an interactive, multimedia tool of cognitive stimulation in Alzheimer's disease. J of Neurol, Neurosurg and Psychiatr. 2006, 77: 1116-1121. 10.1136/jnnp.2005.086074.

Wenisch E, Cantegreil-Kallen I, De Rotrou J, Garrigue P, Moulin F, Batouche F, Richard A, De Sant'anna M, Rigaud AS: Cognitive stimulation intervention for elders with mild cognitive impairment compared with normal aged subjects: preliminary results. Aging Clinical & Experimental Research. 2007, 19 (4): 316-322.

Gates N, Valenzuela MJ: Cognitive exercise and its role in cogitive function in older adults. Current Psychiatry Reports. 2010, 12: 20-27. 10.1007/s11920-009-0085-y.

Verhaeghen P, Marcoen A, Grossens L: Improving memory performances in the aged through mnemonic training: A meta-analytic study. Psychology and Aging. 1993, 7: 242-251.

Zehnder F, Martin M, Algussen M: Memeory training effects in old age as markers of plasticity:a meta-analysis. Restorative Neurology and Neuroscience. 2009, 27: 507-520.

PubMed   Google Scholar  

Mahncke HW, Connor B, Appelmann J, Ahsanuddin : Memory enhancement in healthy older adults using a brain plasticity-based training program: a randomized controlled study. PNAS. 2006, 103 (33): 12523-12528. 10.1073/pnas.0605194103.

Article   CAS   PubMed   PubMed Central   Google Scholar  

Schmiedek F, Bauer C, Lovden M, Brose A, Lindenberger U: Technology in healthy aging. Cognitive enrichment in old age:web-based training programs. Gero Psych. 2010, 23 (2): 59-67.

Jean L, Bergeron ME, Thivierge S, Simard M: Cognitive intervention programs for individuals with mild cognitive impairment: systematic review of the literature. Am J Geriatr Psychiatry. 2010, 18 (4): 281-296. 10.1097/JGP.0b013e3181c37ce9.

Kurz A, Pohl C, Ramsenthaler M, Sorg C: Cognitive rehabilitation in patients with mild cognitive impairment. International Journal of Geriatric Psychiatry. 2009, 24 (2): 163-168. 10.1002/gps.2086.

Valenzuela MJ, Sachdev P: Can cognitive exercise prevent the onset of dementia? Systematic review of randomised clinical trials with longitudinal follow-up. Am J Geriatr Psychiatry. 2009, 17 (3): 179-187. 10.1097/JGP.0b013e3181953b57.

Ash E, Korczyn AD: Is amnestic mild cognitive impairment a useful concept?. Aging Health. 2011, 7 (1): 107-122. 10.2217/ahe.10.90.

Sachdev P, Ganguli M, Petersen C: How can we best categorize cognitive impairment in nondemented older adults?. Am J Geriatr Psychiatry. 2010, 18 (8): 657-10.1097/JGP.0b013e3181e6c2af.

Petersen C: Mild cognitive impairment as a diagnostic entity. J Intern Medicine. 2004, 256: 183-194.

Olazaran J, Muniz R, Reisberg B, Pena-Casanova J, Del Ser T, Cruz-Jentoft AJ, Serrano P, Navarro E: Benefits of cognitive-motor intervention in MCI and mild to moderate Alzheimer disease. Neurology. 2004, 2 (63): 2348-2353.

Troyer A, Murphy K, Anderson N, Moscovitch M, Craik FI: Changing everyday memory behaviour in amnestic mild cognitive impairment: A randomised controlled trial. Neuropsychological Rehabilitation. 2008, 18 (1): 65-88. 10.1080/09602010701409684.

Faul F, Erdfelder E, Buchner A: GPOWER: A general power analysis program. Behavior Research Methods, Instruments, & Computers. 1996, 28: 1-11. 10.3758/BF03203630.

Eckroth-Bucher M, Siberski J: Preserving cognition through an integrated cognitive stimulation and training program. Am Jrnl of Alz Dis & Other Dementias. 2010, 24 (3): 234-245.

Bond G, Wold-Wilets V, Fiedler FE, Burr RL: Computer-aided cognitive training of the aged: a pilot study. Clinical Gerontologist. 2000, 22 (2): 19-42.

Poon P, Hui E, Dai D, Kwok T, Woo J: Cognitive intervention for community-dwelling older persons with memory problems: telemedicine versus face-to-face treatment. International Journal of Geriatric Psychiatry. 2005, 20: 285-286. 10.1002/gps.1282.

Jean L, Simard M, Wiederkehr S, Bergeron ME, Turgeon Y, Hudon C, Tremblay I, van Reekum R: Efficacy of a cognitive training programme for mild cognitive impairment: results of a randomised controlled study. Neuropsychological Rehabilitation. 2010, 20 (3): 377-405. 10.1080/09602010903343012.

Kinsella GJ, Mullay E, Rand E, Ong B, Burton C, Price S, Phillips M, Storey E: Early intervention for mild cognitive imapirment: a randomised controlled trial. J Neurol Neurosurg Psychiatry. 2009, 80: 730-736. 10.1136/jnnp.2008.148346.

Article   CAS   PubMed   Google Scholar  

Londos E, Boschian K, Lindén A, Persson C, Minthon L, Lexell J: Effects of a goal-oriented rehabilitation program in mild cognitive impairment: a pilot study. American Journal of Alzheimer's Disease & Other Dementias. 2008, 23 (2): 177-183. 10.1177/1533317507312622.

Rapp S, Brenes G, Marsh AP: Memory enhancement training for older adults with mild cognitive impairment: a preliminary study. Aging & Mental Health. 2002, 6 (1): 5-11. 10.1080/13607860120101077.

Barnes D, Yaffe K, Belfor N, Jagust W, DeCarli C, Reed B, Kramer J: Computer-based cognitive training for mild cognitive impairment. Results from a pilot randomised, controlled trial. Alzheimer Disease and Associated Disorders. 2009, 23 (3): 205-210. 10.1097/WAD.0b013e31819c6137.

Article   PubMed   PubMed Central   Google Scholar  

Gunther VK, Schafer P, Holzner BJ, Kemmler GW: Long-term improvements in cognitive performance through computer-assisted cognitive training: A pilot study in a residential home for older people. Aging and Mental Health. 2003, 7 (3): 200-206. 10.1080/1360786031000101175.

Petersen RC, Smith GM, Ivnik RJ, Kokmen E, Tangalos EG: Mild cognitive impairment:clinical characterisation and outcome. Archives of Neurology. 1999, 56 (303-308):

Sinforiani E, Banchieri L, Zucchella C, Pacchetti C, Sandrini G: Cognitive rehabilitation in Parkinson's disease. Archives of Gerontology and Geriatrics. 2004, 38 (SUPPL): 387-391.

Marker KR: Marker Software. Heidelberg & Ladenburg. 1986

Papp KV, Walsh SJ, Snyder PJ: Immediate and delayed effects of cognitive interventions in healthy elderly: A review of current literature and future directions. Alzheimer's and Dementia. 2009, 5 (1): 50-60. 10.1016/j.jalz.2008.10.008.

Jean L, Simard M, Wiederkehre S, Bergeron ME, Turgeon Y, Hudon c, Tremblay I, Van Reekum R: Efficacy of a cognitive training programme for mild cognitive impairment: Results of a randomised controlled trial. Neuropsychological Rehabilitation. 2010, 20 (3): 377-405. 10.1080/09602010903343012.

Stott J, Spector A: A review of the effectiveness of memory intervention in mild cognitive impairment. International Psychogeriatrics Association. 2010, 1-13.

Tsai AY, Yang MJ, Lan CF, Chen CS: Evaluation of effect of cognitive intervention programs for the community-dwelling elderly with subjective memory complaints. Int J of Geriatric Psychiatry. 2008, 23: 1172-1174. 10.1002/gps.2050.

Clare L, Woods RT: Cognitive training and cognitive rehabilitation for people with early-stage Alzheimer's disease: A review. Neuropsychological Rehabilitation. 2004, 14 (4): 385-401. 10.1080/09602010443000074.

Clare L, Woods B: Cognitive rehabilitation and cognitive training for early-stage Alzheimer's disease and vascular dementia. Cochrane Database of Systematic Reviews. 2003, (4 Art. No.: CD003260.).

Sitzer DI, Twamley EW, Jeste DV: Cognitive training in Alzheimer's disease: A meta-analysis of the literature. Acta Psychiatrica Scandinavica. 2006, 114 (2): 75-90. 10.1111/j.1600-0447.2006.00789.x.

Cicerone KD, Dahlberg C, Malec JF, Langenbahn DM, Felicetti T, Kneipp S, Ellmo W, Kalmar K, Giacino JT, Harley JP, et al: Evidence-based cognitive rehabilitation: updated review of the literature from 1998 through 2002. Archives of Physical Medicine & Rehabilitation. 2005, 86 (8): 1681-1692. 10.1016/j.apmr.2005.03.024.

Valenzuela MJ, Breakspear M, Sachdev P: Complex mental activity and the aging brain: Molecular, cellular and cortical network mechanisms. Brain Research Reviews. 2007, 56 (1): 198-213. 10.1016/j.brainresrev.2007.07.007.

Thal LJ: Prevention of Alzheimer disease. Alzheimer Disease and Associated Disorders. 2006, 20 (SUPPL. 2): S97-S99.

Galante E, Venturini G, Fiaccadori C: Computer-based cognitive intervention for dementia: Preliminary results of a randomized clinical trial. Giornale Italiano di Medicina del Lavoro ed Ergonomia. 2007, 29 (3 SUPPL. B): B26-B32.

CAS   PubMed   Google Scholar  

Wagner N, Hassanein K, Head M: Computer-use by older adults: a multi-disciplinary review. Computers in human behaviour. 2010, 26 (5): 870-882. 10.1016/j.chb.2010.03.029.

Pre-publication history

The pre-publication history for this paper can be accessed here: http://www.biomedcentral.com/1471-2318/11/55/prepub

Download references

Acknowledgements

The authors would like to thank Ruben Muñiz Schwochert, Javiere Olazaran and Angie Troyer for providing the data requested for inclusion in this review. This work forms part of NG's PhD thesis at the School of Psychiatry, University of New South Wales.

Author information

Authors and affiliations.

School of Psychiatry, University of New South Wales, Randwick, NSW, 2031, Australia

Nicola J Gates, Perminder S Sachdev & Michael Valenzuela

Brain and Aging Research Program, University of New South Wales, Randwick, NSW, 2031, Australia

Perminder S Sachdev & Michael Valenzuela

Regenerative Neuroscience Group, School of Psychiatry, University of New South Wales, Randwick, NSW, 2031, Australia

Michael Valenzuela

Neuropsychiatric Institute, Prince of Wales Hospital, Randwick, NSW, 2031, Australia

Perminder S Sachdev

Exercise Health and Performance Faculty Research Group, Sydney Medical School, The University of Sydney, Lidcombe, NSW, 2141, Australia

Maria A Fiatarone Singh

Hebrew Senior Life, Boston, MA, and Jean Mayer USDA Human Nutrition Research Center on Aging at Tufts University, Boston, 02130, MA, USA

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Nicola J Gates .

Additional information

Competing interests.

The authors declare that they have no competing interests.

Authors' contributions

NG and MV conceived of the study and participated in its design and co-ordination. NG conducted the search, NG and MV completed data extraction; and NG and MV conducted data analysis. NG drafted the manuscript and MV, PS and MFS revised and edited the manuscript. All authors read and approved the final manuscript.

Authors’ original submitted files for images

Below are the links to the authors’ original submitted files for images.

Authors’ original file for figure 1

Rights and permissions.

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License ( http://creativecommons.org/licenses/by/2.0 ), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Reprints and permissions

About this article

Cite this article.

Gates, N.J., Sachdev, P.S., Fiatarone Singh, M.A. et al. Cognitive and memory training in adults at risk of dementia: A Systematic Review. BMC Geriatr 11 , 55 (2011). https://doi.org/10.1186/1471-2318-11-55

Download citation

Received : 25 May 2011

Accepted : 25 September 2011

Published : 25 September 2011

DOI : https://doi.org/10.1186/1471-2318-11-55

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Mild Cognitive Impairment
• Cognitive Training
• Cognitive Rehabilitation
• Memory Strategy
• Cognitive Stimulation

BMC Geriatrics

ISSN: 1471-2318

• Submission enquiries: [email protected]
• General enquiries: [email protected]

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals

Working memory articles from across Nature Portfolio

Working memory is the active and robust retention of multiple bits of information over the time-scale of a few seconds. It is distinguished from short-term memory by the involvement of executive or attentional control that makes the information flexible yet resistant to interference.

Latest Research and Reviews

Volatile working memory representations crystallize with practice

Delay- and choice-related activities that are essential for working-memory performance drift during learning and stabilize only after several days of expert performance.

• Arash Bellafard
• Ghazal Namvar
• Peyman Golshani

Children exhibit superior memory for attended but outdated information compared to adults

Children typically exhibit weaker memory than adults. Here, the authors report a developmental reversal-like phenomenon that children show better memory for attended but outdated information, suggesting underdeveloped memory selection in children.

Alpha neurofeedback training improves visual working memory in healthy individuals

• Wenbin Zhou

Control of working memory by phase–amplitude coupling of human hippocampal neurons

Hippocampal theta–gamma phase–amplitude coupling integrates cognitive control and working memory storage across brain areas in humans.

• Jonathan Daume
• Jan Kamiński
• Ueli Rutishauser

Evidence for a competitive relationship between executive functions and statistical learning

• Felipe Pedraza
• Bence C. Farkas
• Dezső Németh

Chronic phase advances reduces recognition memory and increases vascular cognitive dementia-like impairments in aged mice

• Jennifer A. Liu
• Jacob R. Bumgarner
• Randy J. Nelson

News and Comment

Reply to ‘efficiency and capacity mechanisms can coexist in cognitive training’.

• Claudia C. von Bastian
• Sylvie Belleville
• Tilo Strobach

Efficiency and capacity mechanisms can coexist in cognitive training

• Da-Wei Zhang
• Bruno Sauce

Labels aid visual working memory

• Teresa Schubert

A frequency location to remember

Neuromodulation with specific frequencies at specific brain locations selectively enhances either working memory or long-term memory in older adult humans.

• Jake Rogers

Revisiting mixture models of memory

Probabilistic mixture models have contributed significantly to advancements in visual working memory research in recent decades. In a new paper, Schurgin and colleagues revisit the basic assumptions of mixture models and suggest that we cannot understand memory without first considering perception.

• Blaire Dube
• Julie D. Golomb

Random connections in memory

• Anne-Marike Schiffer

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

The crystallization of memory: Study reveals how practice forms new memory pathways in the brain

A new study led by UCLA Health has shown that repetitive practice not only is helpful in improving skills but also leads to profound changes in the brain's memory pathways.

The research, published in the journal Nature and co-led by Rockefeller University, sought to unravel how the brain's ability to retain and process information, known as working memory, improves through training.

To test this, researchers tasked mice with identifying and recalling a sequence of odors over the course of two weeks. Researchers then tracked neural activity in the animals as they practiced the task by using a novel, custom-built microscope that can image cellular activity in up to 73,000 neurons simultaneously throughout the cortex.

The study revealed a transformation in the working memory circuits located in the secondary motor cortex as the mice repeated the task through time. As the mice were first learning the task, the memory representations were unstable. However, after repeatedly practicing the task, the memory patterns began to solidify or "crystalize," said corresponding author and UCLA Health neurologist Dr. Peyman Golshani.

"If one imagines that each neuron in the brain is sounding a different note, the melody that the brain is generating when it is doing the task was changing from day to day, but then became more and more refined and similar as animals kept practicing the task," Golshani said.

These changes give insights into why performance becomes more accurate and automatic following repetitive practice.

"This insight not only advances our understanding of learning and memory but also has implications for addressing memory-related disorders," Golshani said.

The work was performed by Dr. Arash Bellafard, project scientist at UCLA in close collaboration with Dr. Alipasha Vaziri's group at Rockefeller University.

• Intelligence
• Neuroscience
• Learning Disorders
• Educational Psychology
• Brain Injury
• Memory-prediction framework
• Social cognition
• Limbic system
• Deep brain stimulation
• Asperger syndrome

Story Source:

Materials provided by University of California - Los Angeles Health Sciences . Original written by Will Houston. Note: Content may be edited for style and length.

Journal Reference :

• Arash Bellafard, Ghazal Namvar, Jonathan C. Kao, Alipasha Vaziri, Peyman Golshani. Volatile working memory representations crystallize with practice . Nature , 2024; DOI: 10.1038/s41586-024-07425-w

Cite This Page :

Explore More

• Life Expectancy May Increase by 5 Years by 2050
• Toward a Successful Vaccine for HIV
• Highly Efficient Thermoelectric Materials
• Toward Human Brain Gene Therapy
• Whale Families Learn Each Other's Vocal Style
• AI Can Answer Complex Physics Questions
• Otters Use Tools to Survive a Changing World
• Monogamy in Mice: Newly Evolved Type of Cell
• Sustainable Electronics, Doped With Air
• Male Vs Female Brain Structure

Trending Topics

Strange & offbeat.

February 21, 2024

Why Writing by Hand Is Better for Memory and Learning

Engaging the fine motor system to produce letters by hand has positive effects on learning and memory

By Charlotte Hu

Studies continue to show pluses to writing by hand.

Image Source/Getty Images

Handwriting notes in class might seem like an anachronism as smartphones and other digital technology subsume every aspect of learning across schools and universities. But a steady stream of research continues to suggest that taking notes the traditional way—with pen and paper or even stylus and tablet—is still the best way to learn, especially for young children. And now scientists are finally zeroing in on why.

A recent study in Frontiers in Psychology monitored brain activity in students taking notes and found that those writing by hand had higher levels of electrical activity across a wide range of interconnected brain regions responsible for movement, vision, sensory processing and memory. The findings add to a growing body of evidence that has many experts speaking up about the importance of teaching children to handwrite words and draw pictures.

Differences in Brain Activity

On supporting science journalism.

If you're enjoying this article, consider supporting our award-winning journalism by subscribing . By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

The new research, by Audrey van der Meer and Ruud van der Weel at the Norwegian University of Science and Technology (NTNU), builds on a foundational 2014 study . That work suggested that people taking notes by computer were typing without thinking, says van der Meer , a professor of neuropsychology at NTNU. “It’s very tempting to type down everything that the lecturer is saying,” she says. “It kind of goes in through your ears and comes out through your fingertips, but you don’t process the incoming information.” But when taking notes by hand, it’s often impossible to write everything down; students have to actively pay attention to the incoming information and process it—prioritize it, consolidate it and try to relate it to things they’ve learned before. This conscious action of building onto existing knowledge can make it easier to stay engaged and grasp new concepts .

To understand specific brain activity differences during the two note-taking approaches, the NTNU researchers tweaked the 2014 study’s basic setup. They sewed electrodes into a hairnet with 256 sensors that recorded the brain activity of 36 students as they wrote or typed 15 words from the game Pictionary that were displayed on a screen.

When students wrote the words by hand, the sensors picked up widespread connectivity across many brain regions. Typing, however, led to minimal activity, if any, in the same areas. Handwriting activated connection patterns spanning visual regions, regions that receive and process sensory information and the motor cortex. The latter handles body movement and sensorimotor integration, which helps the brain use environmental inputs to inform a person’s next action.

“When you are typing, the same simple movement of your fingers is involved in producing every letter, whereas when you’re writing by hand, you immediately feel that the bodily feeling of producing A is entirely different from producing a B,” van der Meer says. She notes that children who have learned to read and write by tapping on a digital tablet “often have difficulty distinguishing letters that look a lot like each other or that are mirror images of each other, like the b and the d.”

Reinforcing Memory and Learning Pathways

Sophia Vinci-Booher , an assistant professor of educational neuroscience at Vanderbilt University who was not involved in the new study, says its findings are exciting and consistent with past research. “You can see that in tasks that really lock the motor and sensory systems together, such as in handwriting, there’s this really clear tie between this motor action being accomplished and the visual and conceptual recognition being created,” she says. “As you’re drawing a letter or writing a word, you’re taking this perceptual understanding of something and using your motor system to create it.” That creation is then fed back into the visual system, where it’s processed again—strengthening the connection between an action and the images or words associated with it. It’s similar to imagining something and then creating it: when you materialize something from your imagination (by writing it, drawing it or building it), this reinforces the imagined concept and helps it stick in your memory.

The phenomenon of boosting memory by producing something tangible has been well studied. Previous research has found that when people are asked to write, draw or act out a word that they’re reading, they have to focus more on what they’re doing with the received information. Transferring verbal information to a different form, such as a written format, also involves activating motor programs in the brain to create a specific sequence of hand motions, explains Yadurshana Sivashankar , a cognitive neuroscience graduate student at the University of Waterloo in Ontario who studies movement and memory. But handwriting requires more of the brain’s motor programs than typing. “When you’re writing the word ‘the,’ the actual movements of the hand relate to the structures of the word to some extent,” says Sivashankar, who was not involved in the new study.

For example, participants in a 2021 study by Sivashankar memorized a list of action verbs more accurately if they performed the corresponding action than if they performed an unrelated action or none at all. “Drawing information and enacting information is helpful because you have to think about information and you have to produce something that’s meaningful,” she says. And by transforming the information, you pave and deepen these interconnections across the brain’s vast neural networks, making it “much easier to access that information.”

The Importance of Handwriting Lessons for Kids

Across many contexts, studies have shown that kids appear to learn better when they’re asked to produce letters or other visual items using their fingers and hands in a coordinated way—one that can’t be replicated by clicking a mouse or tapping buttons on a screen or keyboard. Vinci-Booher’s research has also found that the action of handwriting appears to engage different brain regions at different levels than other standard learning experiences, such as reading or observing. Her work has also shown that handwriting improves letter recognition in preschool children, and the effects of learning through writing “last longer than other learning experiences that might engage attention at a similar level,” Vinci-Booher says. Additionally, she thinks it’s possible that engaging the motor system is how children learn how to break “ mirror invariance ” (registering mirror images as identical) and begin to decipher things such as the difference between the lowercase b and p.

Vinci-Booher says the new study opens up bigger questions about the way we learn, such as how brain region connections change over time and when these connections are most important in learning. She and other experts say, however, that the new findings don’t mean technology is a disadvantage in the classroom. Laptops, smartphones and other such devices can be more efficient for writing essays or conducting research and can offer more equitable access to educational resources. Problems occur when people rely on technology too much , Sivashankar says. People are increasingly delegating thought processes to digital devices, an act called “ cognitive offloading ”—using smartphones to remember tasks, taking a photo instead of memorizing information or depending on a GPS to navigate. “It’s helpful, but we think the constant offloading means it’s less work for the brain,” Sivashankar says. “If we’re not actively using these areas, then they are going to deteriorate over time, whether it’s memory or motor skills.”

Van der Meer says some officials in Norway are inching toward implementing completely digital schools . She claims first grade teachers there have told her their incoming students barely know how to hold a pencil now—which suggests they weren’t coloring pictures or assembling puzzles in nursery school. Van der Meer says they’re missing out on opportunities that can help stimulate their growing brains.

“I think there’s a very strong case for engaging children in drawing and handwriting activities, especially in preschool and kindergarten when they’re first learning about letters,” Vinci-Booher says. “There’s something about engaging the fine motor system and production activities that really impacts learning.”

A version of this article entitled “Hands-on” was adapted for inclusion in the May 2024 issue of Scientific American.

How to Improve Memory

Reviewed by Psychology Today Staff

It doesn’t take an extraordinary brain to get smarter about remembering. From techniques used by memory champions to fundamentals like securing enough sleep and maintaining healthy behaviors, just about anyone who wants to learn more efficiently has a variety of tools at their disposal—some of which they have likely already used.

On This Page

• Memory Tricks
• Everyday Memory Boosts

While simply revisiting a newly learned fact, the definition of a word, or some other information can help reinforce someone’s memory for it, additional tools and processes can help make the effort to retain those details more powerful.

• Mnemonic devices are ways of enhancing memory that can involve elaboration—connecting what one is trying to remember to other information in memory—organizing to-be-remembered details more efficiently in memory, and making use of mental visualization. Examples of mnemonics include:

• forming a series of word s into an acronym (such as ROY G BIV, for the colors of the rainbow) or a series of letters into an acrostic (Elephants And Donkeys Got Big Ears, for the notes of each string on a guitar, E-A-D-G-B) • grouping to-be-remembered items together into categories (such as several types of food, when remembering what to buy at the grocery store) • creating a memory palace : visualizing a series of objects, events, or other things appearing in a familiar physical space (such as a room at home), where each one represents something to be remembered; also called the method of loci

• Paying closer attention to details in the moment can make it easier to remember them later. People can learn to focus better; mindfulness techniques may help. Minimizing distractions and avoiding multitasking while learning information could also help with remembering.
• Spacing apart the time spent studying , rather than massing it together, tends to lead to better learning, according to research on the spacing effect . An example of spaced practice would be studying a topic once every day for relatively small blocks of time rather than spending a longer block of time studying on Friday. Accordingly, “cramming”—studying in one long, continuous period— can be an unhelpful study habit.
• Testing memory of learned material , such as a passage of text, can enhance memory for that material—above and beyond re-reading, research indicates. The findings suggest that self-testing can help with learning , whether a person responds to self-generated questions or flashcards related to that information or questions provided by someone else (such as sample test questions in textbooks). Explaining a newly learned concept to oneself or someone else may also help reinforce memory for it.
• Chunking is the combination of to-be-remembered pieces of information, such as numbers or letters, into a smaller number of units (or “chunks”), making them easier to remember. A simple example is the reduction of a phone number into three parts (which one might repeat to oneself in three bursts), though more complex forms of chunking are thought to help account for experts’ superior memory for certain kinds of information (such as chess positions).

Can someone deliberately improve their ability to remember over the long-term? While factors such as well-timed and sufficient sleep and physical activity can aid a neurologically healthy person’s memory ability, the evidence for approaches such as supplements or brain games is often mixed.

In addition to a variety of strategies (such mnemonic devices and others mentioned above) to enhance your memory in the short term , striving to live a healthy and active lifestyle can help preserve memory ability over time. That means engaging in regular mental challenges, exercising routinely, getting enough sleep, and eating well. Reducing stress in daily life may also help to boost memory.

Sleep is thought to play an important role in the consolidation of memories. There is evidence that people who sleep soon after studying new information are more likely to recall it later than those who study it and remain awake. Procedural memories (memory for physical skills, for example) as well as memories for experiences and for new knowledge, seem to benefit from sleep. Consequently, failing to prioritize sleep (or struggling with sleep for other reasons) may mean a missed chance for optimal memory consolidation.

In addition to having longer-term benefits for memory ability, well-timed exercise may immediately boost memory for new information under some conditions. Research has found that moderate-to-high-intensity cardiovascular workout just before or after a learning period enhanced recall for the information learned.

Vegetables, nuts, berries, beans, olive oil, whole grains, fish, and other nutritious foods are elements of the Mediterranean diet and the DASH (Dietary Approaches to Stop Hypertension) Diet, which have been studied for their potential positive long-term effects on brain health. People who, over the course of several years, followed a diet blending elements of both showed reduced risk of Alzheimer’s disease , of which memory loss is one component. The same diet advises limiting consumption of red meat, butter and margarine, cheese, sweets, and fried or fast food.

There is reason to be skeptical about “brain training” programs based on inconsistent evidence of their effectiveness at improving memory or other cognitive abilities. Apps that purport to train the brain often feature tasks used to exercise working memory, with the aim of increasing working memory capacity (which has been linked to intelligence) in order to produce broader cognitive improvements. While working memory training may at least temporarily enhance performance on working memory-related tasks, however, that does not mean the improvement carries over to other mental abilities.

A range of substances, both synthetic and naturally occurring, have been studied for their potential to improve cognitive function, including memory ability. There are certain kinds of medications that can be prescribed to help treat memory loss due to a disease. Supplements proposed to enhance memory in healthy people , however, which have varying degrees of evidence in their favor—often based on small studies—may have a modest impact, if any, on memory.

A few studies have suggested that recalling the past with fondness and gratitude can increase self-control, but a recent meta-analysis challenges this idea.

It remains a matter of scientific debate whether the beta amyloid buildup is the cause of Alzheimer’s or a feature of it. It’s time to look at “out of the clump” fresh approaches.

Researchers developed a method to transform students' writing over 30 years ago. What happened to it?

The evidence strongly points to the perils of long-term use of benzos. This warning is more credible after recent studies have revealed the mechanisms of cognitive impairments.

Has your loved one told you something happened that you’re not sure is true? It could be a false memory.

Older U.S. adults and their families have reason to consider space and place for optimizing older adults' short term memory and attentional needs.

We all grow up with stories about our parents, childhood, and challenges. They form our unique way of looking at life and ourselves, but stories can be distorted. Time to upgrade?

It may require very little daily cannabis consumption to produce long-term neuroprotection in the older brain.

A woman at my gym walks on the treadmill backwards; sometimes quickly, sometimes slowly. After months of watching her and wondering when she might fall, I asked her about it.

Personal Perspective: We are thrilled by even the least coincidental of stories, but do they shape our lives with meaning? Or are they meaningless, random connections?

• Find a Therapist
• Find a Treatment Center
• Find a Psychiatrist
• Find a Support Group
• Find Online Therapy
• United States
• Brooklyn, NY
• Chicago, IL
• Houston, TX
• Los Angeles, CA
• New York, NY
• Portland, OR
• San Diego, CA
• San Francisco, CA
• Seattle, WA
• Washington, DC
• Asperger's
• Bipolar Disorder
• Chronic Pain
• Eating Disorders
• Passive Aggression
• Personality
• Goal Setting
• Positive Psychology
• Stopping Smoking
• Low Sexual Desire
• Relationships
• Child Development
• Self Tests NEW
• Therapy Center
• Diagnosis Dictionary
• Types of Therapy

At any moment, someone’s aggravating behavior or our own bad luck can set us off on an emotional spiral that threatens to derail our entire day. Here’s how we can face our triggers with less reactivity so that we can get on with our lives.

• Emotional Intelligence
• Gaslighting
• Affective Forecasting
• Neuroscience

ORIGINAL RESEARCH article

This article is part of the research topic.

Building the Future of Education Together: Innovation, Complexity, Sustainability, Interdisciplinary Research and Open Science

Developing the Skills for Complex Thinking Research: A Case Study Using Social Robotics to Produce Scientific Papers Provisionally Accepted

• 1 Institute for the Future of Education, Monterrey Institute of Technology and Higher Education (ITESM), Mexico
• 2 University of Cienfuegos, Cuba

The final, formatted version of the article will be published soon.

The development of university students' skills to successfully produce scientific documents has been a recurring topic of study in academia. This paper analyzes the implementation of a training experience using a digital environment mediated by video content materials starring humanoid robots. The research aimed to scale complex thinking and its subcompetencies as a hinge to strengthen basic academic research skills. Students from Colombia, Ecuador, and Mexico committed to preparing a scientific document as part of their professional training participated. A pretest to know their initial level of perception, a posttest to evaluate if there was a change, and a scientific document the students delivered at the end of the training experience comprised the methodology to demonstrate the improvement of their skills. The results indicated students' perceived improvement in the sub-competencies of systemic, creative, scientific, and innovative thinking; however, their perceptions did not align with that of the tutor who reviewed the delivered scientific product. The conclusion was that although the training experience helped strengthen the students' skills, variables that are determinants for a student to develop the knowledge necessary to prepare scientific documents and their derived products remain to be analyzed.

Keywords: higher education, research skills, Educational innovation, complex thinking, scientific thinking, Critical Thinking, Innovative thinking, social robotics

Received: 16 Oct 2023; Accepted: 17 May 2024.

Copyright: © 2024 Lopez-Caudana, George-Reyes and Avello-Martínez. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY) . The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

* Correspondence: Dr. Edgar O. Lopez-Caudana, Institute for the Future of Education, Monterrey Institute of Technology and Higher Education (ITESM), Monterrey, Mexico

People also looked at

Help | Advanced Search

Computer Science > Computation and Language

Title: you only cache once: decoder-decoder architectures for language models.

Abstract: We introduce a decoder-decoder architecture, YOCO, for large language models, which only caches key-value pairs once. It consists of two components, i.e., a cross-decoder stacked upon a self-decoder. The self-decoder efficiently encodes global key-value (KV) caches that are reused by the cross-decoder via cross-attention. The overall model behaves like a decoder-only Transformer, although YOCO only caches once. The design substantially reduces GPU memory demands, yet retains global attention capability. Additionally, the computation flow enables prefilling to early exit without changing the final output, thereby significantly speeding up the prefill stage. Experimental results demonstrate that YOCO achieves favorable performance compared to Transformer in various settings of scaling up model size and number of training tokens. We also extend YOCO to 1M context length with near-perfect needle retrieval accuracy. The profiling results show that YOCO improves inference memory, prefill latency, and throughput by orders of magnitude across context lengths and model sizes. Code is available at this https URL .

Submission history

Access paper:.

• HTML (experimental)
• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

BibTeX formatted citation

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

Harvey Cushing/John Hay Whitney Medical Library

Sharing libraries and collaborating on research papers, ways to share endnote desktop library, 1. endnote   group sharing.

You can share your EndNote library (or groups) through EndNote Desktop

• Pros : Read/write access, easy
• Cons : The sharer can only share (sync) one library, the share-ee can have multiple libraries shared with them, make sure to finish syncing before closing the library, everyone needs to have an EndNote online account  

2. Compress and send an EndNote library (or group) through email or Box

Compress and send your library

How to: File > Compressed Library (.enlx)

• Pros : PDFs can be included in the EndNote library
• Cons : X9.3 isn't backward compatible. If this happens send an RIS file but note, all Groups and Group Sets will be lost, changes made by one user won’t be reflected in other users’ files.  

3. Send your EndNote library embedded within the Word Document you've been editing with EndNote.

Your Word document contains a "Traveling Library" comprising all references cited using EndNote. This enables you to collaborate with your colleagues on the same document without having to have the same EndNote library.

• Pros : Easy
• Cons : No pdfs, metadata incomplete, required that everyone has EndNote Desktop  

Ways to Collaborate on a Paper (Pros and Cons)

1. microsoft word or google docs.

One person makes the edits. One person has access to the EndNote library and has control over the master document.  All others collaborate on the manuscript, editing and marking where a reference goes (use author, year). The person with the master document uses the edited document to add the new references to EndNote and then references into the master document.

• Pros : less prone to error
• Cons : complicated  

2. Google Docs Method

There is no EndNote Cite While You Write tool available for Google Docs.

• Pros : Highly collaborative
• Cons : Complicated

Directions :

• Highlight the references in EndNote
• Drag and drop the references into your Google Doc. This will create an unformatted citation, (it will have curly brackets { }).
• Alternatively, manually insert unformatted citations in the format {first authors surname, year #record number}. To see the record number, in EndNote, right click on the display fields shown in the middle panel and tick Record Number to add it to the display.
• When the manuscript is completed, download the Google Doc as a .docx file
• Open in Word
• Make sure you have a completed EndNote library from which the Word document will pull the references from. If you have people contributing from personal EndNote libraries, the library at the end needs to have all the references in it. If you don't have a copy, ask the person (people) to compress and email their citations.
• Turn on Instant Formatting. If the reference number don't match, Word will prompt you to clarify which reference you mean.

Training and Support

Cushing/whitney medical library.

• Contact your librarian
• WebEx trainings
• EndNote Guide

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• HHS Author Manuscripts

The neurobiological foundation of memory retrieval

Paul w. frankland.

1 Program in Neurosciences & Mental Health, Hospital for Sick Children, Toronto, Ontario, Canada.

2 Institute of Medical Sciences, University of Toronto, Toronto, Ontario, Canada.

3 Department of Psychology, University of Toronto, Toronto, Ontario, Canada.

4 Department of Physiology, University of Toronto, Toronto, Ontario, Canada.

5 Child & Brain Development Program, Canadian Institute for Advanced Research, Toronto, Ontario, Canada.

Sheena A. Josselyn

6 Brain, Mind & Consciousness Program, Canadian Institute for Advanced Research, Toronto, Ontario, Canada.

Stefan Köhler

7 Department of Psychology, University of Western Ontario, London, Ontario, Canada.

8 The Brain and Mind Institute, University of Western Ontario, London, Ontario, Canada.

Memory retrieval involves the interaction between external sensory or internally generated cues and stored memory traces (or engrams) in a process termed ‘ecphory’. While ecphory has been examined in human cognitive neuroscience research, its neurobiological foundation is less understood. To the extent that ecphory involves ‘reawakening’ of engrams, leveraging recently developed technologies that can identify and manipulate engrams in rodents provides a fertile avenue for examining retrieval at the level of neuronal ensembles. Here we evaluate emerging neuroscientific research of this type, using cognitive theory as a guiding principle to organize and interpret initial findings. Our Review highlights the critical interaction between engrams and retrieval cues (environmental or artificial) for memory accessibility and retrieval success. These findings also highlight the intimate relationship between the mechanisms important in forming engrams and those important in their recovery, as captured in the cognitive notion of ‘encoding specificity’. Finally, we identify several questions that currently remain unanswered.

In 1966, Tulving and Pearlstone 1 reported a highly influential finding that profoundly altered the direction of subsequent research on memory in ways that few papers do. Up until this point, almost all experimental research on human memory was concerned with learning or forgetting. The prevalent perspective at the time considered failure in memory performance as the outcome of two possible scenarios. Failure might indicate either that information had not been learned or that it had been learned but subsequently forgotten. However, Tulving and Pearlstone’s work suggested a third possibility. Memory failure could also reflect a problem in retrieval. Specifically, they demonstrated that the same memory could be retrieved successfully with some retrieval cues, but not others ( Fig. 1 ).

Subjects were presented with a series of words. These words were drawn from multiple categories (for example, types of birds, flowers, etc.). In the test phase, subjects were asked to recall as many words as they could from the list (free recall) or from the specific categories (cued recall). The cued recall group performed considerably better than the free recall group across categories, indicating that retrieval cues present at the time of recall determine engram accessibility and subsequent success at remembering.

From this work, Tulving developed an important conceptual distinction between availability versus accessibility of information in memory. According to this view 2 , 3 , some forms of memory failure reflect a lack of availability of pertinent information (i.e., permanent loss), whereas other forms of memory failure reflect temporary problems in accessibility. Phenomenologically, this relates to the common ‘tip of the tongue’ experience, in which one might struggle to recall a familiar name or place while having the strong impression that the information is present. Indeed, often this information subsequently comes to mind. Cues available at retrieval represent perhaps the most critical factor that determines memory accessibility and corresponding success at remembering.

In making this distinction between memory availability versus accessibility, Tulving also recognized 3 earlier work by Richard Semon, a German scientist working at the turn of the twentieth century. Semon 4 first emphasized the role of retrieval cues in remembering and introduced specific terminology to capture this process. Ecphory describes the memory retrieval process, and Semon argued that ecphory reflects the unique interplay between cues and stored memory traces at retrieval. He also coined the term engram’ to refer to such memory traces as biological entities; this may be considered his better-known contribution to the field 5 . Although engrams had not yet been identified empirically, the concept of ecphory became central to the cognitive psychology of memory retrieval 6 .

In the last decade, enormous progress has been made in identifying and manipulating engrams in rodents 7 – 10 . In large part, this progress may be attributed to the development of tools that allow researchers to map engrams to specific neuronal ensembles and manipulate these ensembles using genetically encoded actuators 10 – 15 ( Box 1 ). To date, these approaches have provided evidence for the existence of engrams at the cellular level 7 – 9 , but they may also shed light on the biological basis of memory retrieval 16 , 17 (or, more precisely, ecphory). To the extent that ecphory involves reawakening specific engrams, the ability to identify and manipulate engrams is a prerequisite for gaining mechanistic insights into the retrieval process at the level of neuronal ensembles. Therefore, the recent progress in understanding engrams puts us in position to ask meaningful questions about the neurobiological basis of retrieval. Here we evaluate contemporary neuroscientific research on retrieval at the level of neuronal ensembles using the conceptual framework introduced by Semon and later elaborated by Tulving in his empirical and theoretical work. Although this research also has potentially interesting translational implications, they will not be covered here (but see ref. 18 ).

Approaches for tagging and manipulating engrams in rodents

The allocation strategy takes advantage of the finding that, within a given brain region, eligible excitatory neurons compete for allocation to an engram. This strategy biases which neurons are allocated to an engram by artificially manipulating excitability before a training event. For example, before a training event, a small, random subpopulation of excitatory neurons (purple) is infected with a viral vector expressing a transgene that increases neuronal excitability, such as ChR2 ( Box Fig. a ) 21 , 27 or CREB 20 , 22 , 23 , 108 , 109 . Infected neurons with relatively greater excitability at the time of training are biased for allocation to a resulting engram (red outline). Once allocated, these neurons become both necessary (indispensable) and sufficient (inducing) components of the engram supporting a memory.

In the tagging strategy, neurons that happen to be sufficiently active (that would normally express an activity-dependent immediate early gene) at the time of training are tagged with an actuator (such as an excitatory or inhibitory opsin or chemogenetic construct). To tag active neurons, activity-dependent immediate early gene (IEG) promoters (c-Fos, Arc or others, including synthetic promoters such as E-SARE (enhanced synaptic activity-responsive element) 110 ) are paired with an inducer that ‘opens the tagging window’. Two general types of inducers are used:

• Tetracycline transactivator (tTA)-inducible tagging system. The initial studies 13 , 26 using this approach took advantage of two transgenic mouse lines (but viral vectors can also be used 14 ). In the first transgenic line, tTA (tetracycline-controlled transactivator) is expressed downstream of an IEG promoter. In active cells, neural activity results in tTA expression. However, this process is blocked in the presence of doxycycline (DOX). In second transgenic mouse line, the transgene of interest (depicted as ChR2 in Box Fig. b ) is expressed downstream of a tetracycline response element (TRE). TRE is activated by tTA. Therefore, the absence of DOX opens the tagging window, allowing the transgene of interest to be expressed in active cells.
• Cre recombinase-inducible tagging system. In this system, two transgene cassettes are generally used. In the first, a tamoxifen (TAM)-dependent Cre recombinase (CreER T2 ) is expressed under control of an IEG promoter while in the second, a loxP-flanked STOP signal is placed between a constitutive promoter and the transgene of interest ( Box Fig. c ). In the absence of TAM, the transgene is not expressed. However, in the presence of TAM, Cre recombinase translocates to the nucleus, cleaves the loxP sites, and removes the STOP signal, allowing expression of the transgene. TAM administration opens the tagging window allowing the transgene of interest to be expressed in active cells 12 , 111 .

Manipulating retrieval

Ecphory emphasizes that retrieval reflects interactions between cues, either external sensory or internally generated, and the engram. In other words, memory retrieval can be understood as cue-induced behavioral expression of the engram. It may occur in situations where we intentionally strive to recover a memory in relation to a specific cue (for example, trying to remember where we initially encountered a person we just again met). In other situations, cues may spontaneously trigger memory retrieval (for example, seeing a picture of Paris and remembering a recent visit there).

Contemporary engram studies have examined ecphory in three ways. The first type of experiment asked whether it is possible to prevent ecphory in the presence of external sensory retrieval cues ( Fig. 2a ). For instance, Tanaka and colleagues 19 used a tetracycline-based system (TetTag) to label a contextual fear memory engram in mice, such that CA1 neuronal ensembles that were active during conditioning expressed an inhibitory opsin (ArchT). When subsequently placed back in the training context, mice typically freeze, indicating they recognized that this was the place where the footshock was previously administered. Critically, optogenetic inhibition of the ArchT-tagged neuronal ensemble during this test session reduced conditioned freezing levels (indicating impairment in memory retrieval). Of particular relevance, from the perspective of ecphory, is that the freezing behavior was context-specific (i.e., cue-specific). When a non-overlapping neuronal ensemble tagged in a different context (context B) was silenced during contextual fear testing in context A, mice froze, indicating that this intervention did not interfere with retrieval of the context A fear memory. Similar disruption of cue-induced retrieval by silencing corresponding engrams was observed across a variety of experimental conditions. These include silencing other brain regions (for example, the amygdala 20 – 22 and insular cortex 23 ), in tasks other than fear conditioning (for example, cocaine-cue memory 24 ), as well as using alternate genetic ensemble tagging systems (for example, cre-inducible systems 11 , 12 , 25 ).

a , In this experiment 19 , neuronal ensembles in the CA1 region of the hippocampus were tagged with the inhibitory opsin, ArchT, during contextual fear conditioning (left). When placed back into the training context (i.e., the retrieval cue), mice froze (middle). However, optogenetic inhibition of the tagged ensemble during this test reduced freezing levels (right), indicating that engram silencing can prevent ecphory even in the presence of natural retrieval cues. b , In this experiment 26 , neuronal ensembles in the DG region of the hippocampus were tagged with the excitatory opsin, ChR2, during contextual fear conditioning (left). When placed into a distinct context, mice did not freeze (middle). However, optogenetic activation of the tagged ensemble during this test induced freezing (right), indicating that engram activation, in the absence of natural retrieval cues, can induce ecphory.

The second type of experiment asked the converse question: is it possible to induce memory expression in the absence of sensory retrieval cues via direct stimulation of a tagged engram ( Fig. 2b )? For instance, Liu and colleagues 26 used a similar TetTag approach to express an excitatory opsin (ChR2) in neuronal ensembles that were active during contextual fear conditioning. Following conditioning, placing mice in a context distinct from the training context resulted in little freezing behavior. However, direct photostimulation of the ChR2-tagged neuronal ensemble in the dentate gyrus (DG) induced freezing. Subsequent studies generalized these findings across experimental conditions 11 , 25 and in other brain regions (including the lateral amygdala (LA) 27 – 30 , basolateral amygdala (BLA) 31 and retrosplenial cortex 32 ). Together, these types of experiments indicate it is possible to bypass the requirement for natural retrieval cues in ecphory and to induce memory expression via direct stimulation of the putative engram. One interpretation is that stimulation reflects a reinstatement of an otherwise natural cue.

The first two types of studies used experience-dependent tagging approaches to label neurons that were endogenously active at the time of an event, and then used artificial means (for example, photostimulation) to either block or elicit ecphory. This begs the question of whether the opposite is possible: to create an engram by artificial means and then probe ecphory using natural cues. This question has been addressed in the third type of study considered here ( Fig. 3 ). In this study 33 , photostimulation of a specific olfactory glomerulus (M72) was paired with photostimulation of specific projections that mediate aversion (from the lateral habenula to the ventral tegmental area (VTA)) to create an artificial engram. When mice were subsequently presented with a real M72-activating odorant (acetophenone), they exhibited conditioned avoidance, even though they had not encountered this odor previously. If, instead, M72 activation was paired with photostimulation of reward-mediating projections (laterodorsal tegmental nucleus → VTA), mice subsequently approached, rather than avoided, the M72 odorant, acetophenone. Retrieval of these artificially generated memories and real odor memories (in which acetophenone was actually paired with shock) engaged similar neural circuits, and suppressing neuronal activity in the BLA prevented expression of both artificial and real memories. Three aspects of this work illustrate nicely the tight interplay between engrams and retrieval cues, as initially suggested by Semon. First, artificial engram expression was demonstrated via presentation of a natural external sensory retrieval cue. Second, memory expression reflected the predicted content of the stored information (i.e., mice either approached or avoided acetophenone, depending on which VTA inputs, rewarding or aversive, were stimulated during the training phase). Third, behavioral responding was restricted to the trained cue and did not occur in the presence of unrelated cues.

a , In these experiments 33 , mice formed either a real (top) or an artificial (bottom) odor aversion memory. For the real odor memory, an odor (acetophenone; green) was paired with shock during training. When mice were subsequently presented with the conditioned odor (acetophenone) or a distinct odor (carvone; orange), mice exhibited conditioned aversion to acetophenone. For the artificial odor memory, photostimulation of a specific olfactory glomerulus (M72) was paired with photostimulation of lateral habenula inputs into the VTA. When mice were subsequently tested, they avoided the M72 odorant acetophenone (green), preferring to spend time on the carvone (non-M72 odorant; orange) side of the apparatus. b , In these experiments 33 , mice formed either a real (top) or an artificial (bottom) odor attraction memory. For the real odor memory, an odor (acetophenone; green) was paired with food during training. When mice were subsequently presented with the conditioned odor (acetophenone) or a distinct odor (carvone; orange), mice exhibited conditioned attraction to acetophenone. For the artificial odor memory, photostimulation of the M72 olfactory glomerulus was paired with photostimulation of laterodorsal tegmental nucleus inputs into the VTA. When mice were subsequently tested, they approached (rather than avoided) the M72 odorant acetophenone (green), even though they had never had never encountered this odor previously.

Accessibility of engrams

The studies reviewed so far indicate that it is possible to both disrupt and to mimic ecphory by directly manipulating the activity of neuronal ensembles that were active during encoding. However, they do not address Tulving’s distinction between engram accessibility versus availability. Another category of studies speaks to this distinction, aiming to recover apparently ‘lost’ memories via direct optogenetic stimulation of the tagged engram. By doing so, these studies shed light on the biological mechanisms that distinguish whether a memory can be accessed in principle or not (i.e., when it is unavailable).

In one experiment, Ryan and colleagues 30 tagged neuronal ensembles in either the DG or CA1 region of the hippocampus that were activated during contextual fear conditioning. Immediately following training, mice were treated with the protein synthesis inhibitor anisomycin and were tested 1 day later by returning mice to the training context. As expected, protein synthesis inhibition impaired consolidation and prevented subsequent memory expression. Despite this apparent amnesia in the presence of natural retrieval cues, however, optogenetic reactivation of the tagged neuronal ensemble enabled memory recovery 30 .

Similar recovery from amnesia has been observed across a range of conditions. For instance, following post-training protein synthesis inhibition, artificial engram reactivation in the DG or LA allows for recovery of place aversion or tone fear memories, respectively 30 , 34 . Moreover, memory recovery is not limited to amnestic states produced by protein synthesis inhibition during the consolidation period. Protein synthesis inhibition following natural memory retrieval blocks reconsolidation 35 , 36 , and this lost memory can be recovered via artificial engram reactivation 30 . Memory recovery has also been observed from other amnestic states, including in mouse models for studying Alzheimer’s disease 37 , 38 , infantile amnesia that naturally occurs in early development 39 , and following natural forgetting of social memories 40 .

These results suggest that the underlying engram corresponding to the presumably forgotten event is not completely erased or, using Tulving’s terminology, unavailable. Rather, these engrams exist in otherwise inaccessible states, in which natural retrieval cues (such as exposure to the training context) typically are not sufficient to induce successful ecphory and resulting memory expression. Engrams in this state have been termed ‘silent’ 37 . This is distinct from the notion of latent engrams introduced by Semon, which are both available and accessible through natural cues in principle, only not being accessed in the moment. By contrast, the silent engram is an in-between state: it is available, but nonetheless inaccessible by any natural means. Recent work shows that during engram formation, there is a specific increase in synapses between ‘engram cells’ 30 , 41 , 42 . Maintaining these enhanced synaptic connections may be key to their later accessibility, as evidence suggests that weakening synaptic connections among the neurons of the critical ensembles and, additionally, between these ensembles and downstream regions, is associated with engram silencing 30 , 34 , 37 . Direct photostimulation of the silent engram may temporarily reinstate these weakened connections, leading to memory recovery.

While photostimulation of silent engrams induces memory expression, memory recovery is only transient: freezing behavior is typically only observed during photostimulation 30 , 34 , 37 – 40 . The absence of memory expression in the light-off epochs suggests that the engram remained inaccessible by natural cues. Might interventions that permanently reinstate connectivity shift an engram from a silent state back into a latent state, where it is available and accessible through natural cues? A number of strategies have been used to address this question. For instance, spine density is reduced on DG and CA1 neurons in mouse models for Alzheimer’s disease. High-frequency photostimulation of perforant path afferents (i.e., ‘opto-LTP’) restores spine density on these engram cells, as well as their connectivity to downstream targets (for example, in CA3 and BLA). Critically, in these experiments, presentation of natural cues (i.e., the training context) was now sufficient to induce memory expression in tests performed several days later, suggesting that the opto-LTP intervention had successfully transformed the engram from a silent to latent state 37 . Similarly, overexpression of a dominant active form of PAK1 in experience-tagged CA1 neurons restores spine density and allows memories lost through protein synthesis inhibition to be recovered by natural cues 34 . In related work, Nabavi and colleagues 43 demonstrated that it was possible to modulate engram accessibility by manipulating the strength of synaptic inputs to the LA using opto-LTP (long-term potentiation) and opto-LTD (long-term depression) protocols.

The general picture emerging from this work is that engrams can differ in their degree of accessibility ( Fig. 4 ) and that changes in accessibility reflect underlying changes in synaptic organization. Silent engrams are unique in that they can only be accessed by artificial means. The silent state may be transitional and mark the boundary between lack of engram accessibility and availability.

Engrams exist in a dormant state (where natural retrieval cues induce engram activation and successful retrieval), a silent state (where only direct optogenetic engram activation induces successful retrieval) and an unavailable state (where all information has been lost, and the memory is inaccessible regardless of the nature of access attempts). Transitions from dormant → silent→ unavailable likely reflect forgetting mechanisms (for example, weakening and loss of synaptic connectivity among engram cells or the addition of new connectivity as a consequence of neurogenesis). LTD, long-term depression.

Above we discussed the fact that some seemingly lost memories may simply be inaccessible by natural cues. Are some memories entirely unavailable? This is a difficult, if not impossible, question to answer. To the extent that any testing involves exploration with a finite number of cues, it is always a possibility that successful memory recovery could be achieved with cues that were not tested 44 . Similarly, failure to recover memories with optogenetic stimulation of tagged ensembles might simply reflect failure to test all stimulation protocols. Methods allowing for unambiguous labeling of specific engrams might one day offer researchers the unique opportunity to determine whether an engram has completely disappeared and is truly unavailable. While there are indeed techniques that allow permanent labeling of different components of engrams (for example, at the neuronal ensemble level 45 or synapse level 41 ), it is not clear at what point one could conclude that the absence of a marker indicates that the engram is completely gone. There might always be other markers that could point to remnants of the engram.

That being said, a large body of research shows that forgetting curves have canonical forms that ultimately approach zero (performance level) across whichever behavioral assessments are employed. Recent studies have identified a variety of active forgetting mechanisms at the neurobiological level, including dopamine-initiated signaling cascades, receptor trafficking and hippocampal neurogenesis, all of which could lead to erosion of the engram 46 – 48 . While this line of research is still in its infancy, this class of mechanisms may be of the kind that leads to silencing, ultimately rendering the engram unavailable over the course of forgetting, regardless of the nature of access attempts.

Retrieval as neuronal reinstatement

Recognition of the important distinction between accessibility and availability in cognitive psychology, which began with Tulving and Pearlstone’s findings 1 , led to critical insights on the relationship between encoding and retrieval. To understand what constitutes an effective retrieval cue, it is necessary to consider how the engram was initially formed. Specifically, Tulving and Thomson 49 hypothesized that an engram is shaped by environmental features and internal cognitive or affective states during encoding. In turn, they argued, retrieval cues can only be successful to the extent that they overlap with these environmental features and internal states: that is, the greater the match between encoding and retrieval states, the higher the probability of retrieval success, a principle they termed ‘encoding specificity’. At the behavioral level, evidence in human and nonhuman species suggests that reinstatement of encoding context at the time of retrieval boosts recovery of information acquired in this context 50 – 52 . In fear conditioning, such context specificity provides the organism with adaptive flexibility, ensuring that expression of conditioned fear is usually limited to the training context (or very similar contexts) 53 , 54 .

In contemporary functional neuroimaging and recording studies in humans, the encoding specificity principle has been linked to neuronal reinstatement. Research asking to what extent neural activity patterns at encoding and retrieval overlap provides evidence for spatial and temporal forms of reinstatement that supports this principle 55 – 62 . Moreover, such studies reveal that the extent of this overlap impacts success and phenomenological attributes of retrieval. For instance, in visual cortex, increasing activation overlap predicts memory vividness during retrieval 63 . Interestingly, retrieval success may depend on concurrent hippocampal engagement, not only during encoding 64 but also during retrieval 65 , 66 , with the latter perhaps reflecting a pivotal role of the hippocampus in pattern completion. The importance of neuronal reinstatement for context-specific retrieval has been demonstrated in work showing that its behavioral benefits are most pronounced when encoding and retrieval context match 67 .

The encoding specificity principle can also be evaluated in rodent studies in which cue-induced reactivation of neuronal ensembles active at encoding is examined at the cellular level. Initial work took advantage of a method that images the subcellular location of mRNA for the immediate early gene Arc , catFISH (cellular compartment analysis of temporal activity by fluorescent in situ hybridization), as a way to identify active neurons at two distinct time points. In this experiment 68 , rats were exposed consecutively to either identical (‘AA’ condition) or different (‘AB’ condition) environments, and neuronal ensembles activated by each exposure were assessed. In hippocampal CA1, higher levels of overlap in the AA, compared to the AB, condition suggested that retrieval re-engaged the neuronal ensemble active during initial encoding. While this study did not examine a behavioral readout of memory, subsequent studies linked behavioral expression of memory at retrieval to reactivation of the ensemble active at encoding using ensemble tagging approaches 12 – 14 , 25 , 69 – 72 . For instance, Reijmers and colleagues 13 trained mice in a tone fear conditioning paradigm. Subsequent replacement in the training context reactivated neurons in the basal amygdala at above chance rates. Crucially, the rate of reactivation predicted memory strength, supporting the idea that greater similarity between encoding and retrieval states is associated with greater probability of retrieval success 73 .

In agreement with results examining context specificity in human neuroimaging 67 , studies in rodents reveal that neuronal ensembles activated at retrieval show context specificity related to behavior 25 , 74 . In one study 74 , a tone was paired with footshock in context A during training. Rats were subsequently given extinction training in context B, and then the tone conditioned stimulus (CS) was presented both in the extinction context (context B) and a third, distinct context (context C). Consistent with the idea that extinction is context-specific, rats froze in context C but not in context B (the extinction context) in these tests. At the neuronal level, presentation of the same tone CSs activated distinct populations of neurons in the B and C contexts. Moreover, activation of these different neuronal populations was critical for context-specific expression of extinction 25 .

Given that natural retrieval cues reactivate neural ensembles active at encoding and that the rate of reactivation relates to the strength of memory expression, we can ask whether the same holds for artificially induced memory retrieval. Recent studies 30 , 34 have addressed this question. In these studies, during contextual fear conditioning, cells active in the CA3 and BLA were tagged. Posttraining, mice were administered a protein synthesis inhibitor to silence these engrams. As expected of a silent engram, no freezing was observed when the mice were placed back in the training context. However, optogenetic reactivation of the tagged DG cells produced freezing, and reactivation efficiency (i.e., the extent to which photostimulation induced reactivation of tagged encoding cells) predicted the strength of the artificially retrieved memory (i.e., freezing levels).

While many studies show that artificial reactivation of engrams induces memory expression, typically this expression is weaker than that evoked by natural cues. This finding is in agreement with the encoding specificity principle because it is unlikely that optogenetic stimulation fully recapitulates the state of the organism and the corresponding patterns of neural activity that occurred during encoding. While the local spatial features of activity patterns are preserved by optogenetic stimulation, temporal features are not faithfully reproduced. The development of holographic photostimulation approaches (that preserve both spatial and temporal patterning) may overcome this limitation of current optogenetic techniques 75 – 77 . In the future, closed-loop optogenetic systems could allow the recording and subsequent holographic reproduction of an endogenous ecphoric event 78 , 79 .

Although artificial engram manipulations are typically focal in nature, their effects may be more widespread. Experiences are encoded in hippocampal-cortical networks, and according to many contemporary accounts, the hippocampus plays a pivotal role both in the formation of memory as well as its recovery. At retrieval, the hippocampus is thought to reinstate patterns of activity in the cortex that were present at encoding 80 – 83 . Tanaka and colleagues 19 tested this idea by tagging CA1 neuronal ensembles that were active during contextual fear conditioning. Silencing these tagged hippocampal cells during retrieval impaired memory expression and, critically, reduced reactivation of tagged cortical ensembles.

Conversely, activation, rather than inhibition, of tagged hippocampal neurons reinstates patterns of cortical activity present at encoding. For instance, Guskjolen and colleagues 39 trained infant mice in contextual fear conditioning, tagging active ‘encoding’ ensembles with ChR2. When these mice were tested at later time points, they exhibited pronounced forgetting, a phenomenon resembling infantile amnesia in humans 84 . However, photostimulation of ChR2-tagged neurons in the DG induced memory recovery and reactivation of CA3, CA1 and cortical neurons that were tagged during training.

These types of findings support the idea that some engrams are distributed, spanning neuronal ensembles across subcortical and cortical brain regions 85 . Within this distributed network, each region may carry unique information about the encoded episode (for example, sensory, affective, spatial information), and the route by which network activation is triggered likely impacts phenomenological aspects of memory retrieval. The finding that activation of the hippocampus is essential for reinstating patterns of activity in the cortex that occurred during encoding (as also suggested by human neuroimaging studies 65 , 66 ) additionally supports the view that the hippocampus is a critical hub within these distributed networks. However, it is unlikely that this region is the only hub with a critical role in reinstatement of neuronal states during retrieval 32 , 86 . Moreover, which regions serve as hubs likely changes over time, reflecting ongoing processes that modify the engram after initial memory formation, including consolidation and transformation 87 , 88 .

Equivalency

Artificially reactivating a naturally formed engram induces memory expression. But is ecphory induced by artificial means equivalent to natural ecphory? Next, we highlight four aspects of equivalency between artificially and naturally induced memories.

First, a naturally retrieved memory can serve as a CS for new learning 89 . A study by Ramirez and colleagues 71 tested whether an artificially retrieved memory can similarly support new learning. In this experiment, neuronal ensembles activated by placing a mouse in a neutral context (context A) were tagged with ChR2. One day later, mice were foot-shocked in a second context (context B) while the tagged neuronal ensemble in the DG (corresponding to context A) was simultaneously reactivated. In subsequent testing, mice froze in context A (but not in a dissimilar context, C), even though context A had never been paired with footshock. A study by Ohkawa and colleagues 90 went further. They used similar approaches to separately tag hippocampal and amygdala ensembles corresponding to context exposure (CS) and shock exposure (unconditioned stimulus, US), respectively. To create an artificial association between these ensembles corresponding to otherwise discontiguous events, the tagged CS and US ensembles were synchronously reactivated in the mouse’s home cage. Remarkably, when later placed in the original context, mice now froze even though they had never received a shock in this context.

Second, naturally retrieved memories extinguish. Repeated CS presentations in the absence of US lead to reduced conditioned responding. Khalaf and colleagues 70 asked whether artificially retrieved fear memories similarly extinguish. To do this, they tagged hippocampal ensembles that were activated when mice were placed in a training context that had previously been paired with footshock. Repeated exposure to this training context led to a reduction in freezing behavior (i.e., extinction). However, reactivating the tagged hippocampal ensembles during extinction training accelerated extinction. Conversely, silencing this same population during extinction training slowed extinction. Recently, a related study tagged dorsal hippocampal ensembles during contextual fear conditioning. They then found that repeated, artificially induced retrieval, even in the absence of exposure to the training context, induced extinction of the contextual fear memory 91 .

Third, naturally retrieved memories reconsolidate. Retrieval destabilizes engrams, and protein synthesis is necessary for their restabilization (a process termed reconsolidation). Kim and colleagues 28 asked whether a reconsolidation-like process occurs following artificially induced memory retrieval. In their experiment, CREB-overexpressing neurons in the LA were allocated to a tone fear memory during training. Artificially reactivating this allocated ensemble induced memory expression. However, pharmacological blockade of protein synthesis following artificial induction of ecphory impaired reconsolidation: when subsequently presented with the tone (i.e., the natural cue), mice treated with the protein synthesis inhibitor showed memory disruption. These results indicate that either artificial or natural retrieval destabilizes engrams, leading to the requirement for protein synthesis for their subsequent restabilization.

Fourth, naturally retrieved memories are subject to interference. If similar events are encountered either before or following the event in question, recovery of this target event can be compromised. That is, the ‘wrong’ (i.e., non-target) event or a merged event that combines the target and a similar lures could be recovered 92 . A similar phenomenon was observed following artificially induced retrieval in mice. Garner and colleagues 93 tagged neuronal ensembles activated by exposure to a neutral context (context A) with the excitatory designer receptor exclusively activated by designer drug (DREADD) hM3Di. Mice were subsequently trained in a second context (context B) and tested 24 h later in the same context (context B). Chemogenetic activation of the context A ensemble while testing in context B reduced freezing levels, suggesting that reactivating the ‘wrong’ event interfered with natural cue-induced retrieval of the context A memory.

Retrieval over time: future challenges

This Review highlights the considerable progress made in gaining mechanistic insight into the process of memory retrieval at the biological level. This progress has been enabled by the development of new technologies that allow engrams to be visualized and manipulated in rodents at the level of neuronal ensembles. Combining this increased understanding of engrams with the cognitive theory developed by Endel Tulving 2 permitted us to interpret contemporary research findings with respect to two major themes. First, when viewed in total, neurobiological findings support the cognitive theory that engram accessibility and memory retrieval success critically depend on interactions between engrams and retrieval cues (environmental or artificial). Second, the data also support the close ties between forming of engrams and their recovery, as captured by the notion of encoding specificity. However, the neurobiological study of retrieval is still in its infancy, and many important questions remain unanswered. We emphasize some of the most pressing issues in this remaining section.

Broadly speaking there is a dearth of knowledge as to how processes operating on engrams after their formation influence mechanisms of retrieval. Post-formation changes to the engram can be considered at two levels 87 . First, an engram for an individual episode or event changes over time. Second, multiple engrams (of distinct events or for the same re-encoded event) may interact. We assume that both types of change, which are likely not independent and are often considered together under the broad umbrella of systems consolidation, affect mechanisms of retrieval.

Psychological research suggests that forgetting is not indiscriminate and typically preserves gist over detail in the retention of events (for example, ref. 94 ). It has been argued that this property is adaptive, with gist being particularly important when using memory to guide future behavior and make related predictions 95 . Currently, it is unclear how these dynamics and resulting changes in engram organization affect the neural mechanisms of retrieval. A shift toward more gist-like representation likely occurs hand-in-hand with large-scale shifts in network engagement during retrieval. For example, it has been proposed that retrieval of a gist-based representation (lacking episodic detail) may increasingly engage cortical regions over time, and, furthermore, hippocampal integrity may not be required for its retrieval 96 . At the level of neuronal ensembles, this shift toward more gist-like representation may involve partial silencing of hippocampal engrams. One recent study in mice 97 labeled cells active during contextual fear conditioning in DG and medial prefrontal cortex (mPFC). When placed back in the context 1 day after training, only the DG engram was reactivated (whereas the mPFC engram was not). However, when tested 12 days after training, the mPFC engram, but not the DG engram, was engaged. Nonetheless, optogenetic stimulation of the DG engram (at the remote time point) or the mPFC engram (at the recent time point), respectively, induced artificial memory expression in an alternate context 97 . These changes can be understood as region-specific shifts in engram accessibility (rather than availability) 98 , which may go hand-in-hand with changes in the specificity of the memory expressed in behavior.

Beyond the fate of individual engrams, interactions between engrams may also influence subsequent memory retrieval. Indeed, there is a rich cognitive neuroscience literature focusing on the extraction of regularities across multiple experiences 99 and the resulting changes in network engagement during retrieval. Data addressing this question at the level of neuronal ensembles, however, are only beginning to emerge. An initial study by Rashid and colleagues 21 revealed that the engrams underlying two events experienced within a short period of time (<6 h) engage overlapping engrams and serve to link the two events, such that recall of one event produces recall of the other. In contrast, engrams supporting the same two events experienced with a longer intervening time (24 h) engage non-overlapping neural ensembles, and these events are remembered separately. Moreover, recalling an older event in the hours before experiencing a new event also links the two memories. Although these findings were initially reported for auditory fear memories and neural ensembles in the LA, other groups reported similar findings in the hippocampus supporting two context memories 100 and a conditioned fear and conditioned taste aversion memory in the LA 101 . These findings provide evidence supporting the notion that once formed, engrams do not persist in isolation. However, as of yet the findings do not offer any insight that directly speaks to consequences for mechanisms engaged during retrieval.

One outcome of the extraction of regularities across multiple experiences is the development of schemas 102 . Schemas have received much attention in psychological research on retrieval, but have only recently been studied using neurobiological methods, albeit with promising initial results 103 , 104 . How schemas are organized at the level of neuronal ensembles, however, remains uncharted territory. It has been argued that the availability of a schema qualitatively changes the retrieval process; rather than directly accessing an engram, retrieval involves the reconstruction of a specific episode based on schema knowledge derived from multiple experiences 105 . It is difficult to determine, in particular for remote memories, the extent to which neuronal activity during retrieval reflects such reconstruction vs true engram reactivation 106 , 107 .

Here we have reviewed the current state of knowledge on the mechanisms of memory retrieval at the level of neuronal ensembles. Although recent progress in developing techniques for identifying and manipulating engrams at the level of neuronal ensembles has increased our understanding of engrams in the rodent brain, our understanding of the neurobiological underpinnings of retrieval remains rudimentary Guided by cognitive theories of ecphory, here we integrated and interpreted the findings of several studies taking advantage of the ability to tag and manipulate engrams. We hope this will spur further neuroscientific research into mechanisms underlying retrieval.

Acknowledgements

We thank A.Ramsaran and A.Park for drawing the figures, and we thank T. Ryan for comments on an earlier draft of this manuscript. This work was supported by Canadian Institutes of Health Research grants to P.W.F. (FDN-143227) and S.A.J. (FDN-388455) and a Natural Sciences and Engineering Research Council Discovery grant to S.K. (RGPIN-5770).

Competing interests

The authors declare no competing interests.

Peer review information Nature Neuroscience thanks Stephen Maren and Steve Ramirez for their contribution to the peer review of this work.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.