Memory reactivation generates new, adaptive behaviours that reach beyond direct experience

Abstract

Periods of rest and sleep help us find hidden solutions to new problems and infer unobserved relationships between discrete events. However, the mechanisms that formulate these new, adaptive behavioural strategies remain unclear. One possibility is that memory reactivation during periods of rest and sleep has the capacity to generate new knowledge that extends beyond direct experience. Here, we test this hypothesis using a pre-registered study design that includes a rich behavioural paradigm in humans. We use contextual Targeted Memory Reactivation (TMR) to causally manipulate memory reactivation during awake rest. We demonstrate that TMR during rest enhances performance on associative memory tests, with improved discovery of new, non-directly trained associations, and no change observed for directly trained associations. Our findings suggest that memory reactivation during awake rest plays a critical role in extracting new, unobserved associations to support adaptive behavioural strategies such as inference.

Fig. 1

Behavioural task design and learning performance. (A) Study timeline. The study was split into the learning phase, rest + TMR phase, and testing phase. During the learning phase, the contextual background music was played during trials on each of the respective maps. During rest + TMR, the contextual music for the TMR map was played to bias memory reactivation towards that map (fully counterbalanced across participants). During the rest + TMR phase, participants alternated between 10 min eyes closed rest and 20 min jigsaw puzzle, for a total of 90 min. After the rest + TMR phase, participants completed the testing phase during which they underwent a series of tests to assess the effect of manipulating memory reactivation using TMR on behaviour. (B-C) Task schematic. Participants learned two maps of associations: (B) map 1: a pink house; (C) map 2: a green house; fully counterbalanced across participants. Each map consisted of 11 nodes with each node (room) containing a unique scene and cue (symbol). During learning, each map was paired with a different contextual background music track (café or jungle music; fully counterbalanced across participants). (D) Example trial during the learning phase. Participants learned associations between rooms (scene + cue) and adjacent cues on the map, via sequences of three-alternative-forced-choice trials, always traversing the map in a clockwise direction. At the start of each block of trials for each map, participants were shown the current house (pink or green). On each trial, participants started at a node with an arrow pointing right. Participants were required to choose the symbol in the adjacent room from three options provided at the bottom of the screen, before moving to the adjacent room and receiving feedback. (E) Learning accuracy during first learning session. Participants alternated between blocks of training for each map (example shown by pink/green colours; order counterbalanced across participants). Final mean accuracy: 69.77% (SEM = 1.93%). (F) Learning accuracy in second learning session. N = 38 participants did not reach criterion on the first learning session and proceeded to the second learning session, where participants alternated between shorter blocks of training for each map (example shown by pink/green colours; order counterbalanced across participants). Mean accuracy across the final block of learning for all participants: 81.68% (SEM = 0.61%). (G) Learning accuracy across the final learning blocks of the second learning session, for maps that subsequently became the TMR map (blue) and the no-TMR map (red). (H) In the final learning block there was no significant difference in learning accuracy between maps later assigned to TMR and no-TMR across all participants (n = 40, p = 0.604). Left: raw data points for no-TMR map (red; left) and TMR map (blue; right); each data point is mean accuracy for one participant; black dot, mean; black ticks ± SEM. Right: difference in mean percentage accuracy between no-TMR and TMR maps shown using bootstrap-coupled estimation (DABEST) plots. Effect size for the difference between no-TMR and TMR maps was computed from 10,000 bias-corrected bootstrapped resamples: black dots, mean; black ticks, 95% confidence interval; filled-curve, sampling-error distribution.

Fig. 2

Schematics of the 8 tasks used during the post-rest memory tests. The 8 test tasks (A-G) are shown in the order in which they were completed by participants. (A) Scene-scene: participants were shown a scene and asked to identify the scene in the neighbouring node. (B) Scene-cue: participants were shown a scene without its paired cue and asked to identify the missing cue. (C) Cue-cue: participants were shown a cue without its paired scene and asked to identify the cue from the neighbouring node. (D) One node step: as in the learning task, participants were shown a scene with its paired cue and asked to identify the cue from the neighbouring node. (E) Map 1/map 2: participants were shown a cue and asked whether it belongs to the pink or green house. (F) Same/different: participants were shown two cues and asked if they belong to the same map or different maps. (G) Two node steps: participants were shown a room (scene + cue) and were asked to identify the cue two node steps to the right. (H) Shortest route: while trial sequences in the learning phase were delivered in a clockwise direction only, here participants were shown a room (scene + cue) and were required to indicate the shortest route to a probe cue, positioned two nodes away, by selecting either clockwise (right) or anticlockwise (left) directed travel around the underlying map. This task was designed to assess participants’ ability to abstract out the underlying ring structure of the memory maps, before using this knowledge to execute efficient navigation.

Fig. 3

Performance on post-rest memory tests reveals TMR effect for non-directly trained associative tests. (A) Schematic showing battery of tests included in the test phase to assess the effect of TMR on different aspects of memory. Associative memory tests shown in purple (scene-scene, scene-cue, cue-cue, one node step, and two node step inference) test knowledge of within-map associations. Contextual memory tests shown in orange (map1/map2, same/different) test between-map knowledge. The navigation memory test shown in blue-grey (shortest route) tests participants’ ability to navigate the underlying memory maps. (B-I ) Left: raw data points for no-TMR group (red; left) and TMR group (blue; right) (B-E, G-H), or for two different types of task (green, left; purple, right) (F, I). Each data point is mean accuracy for one participant; black dot, mean; black ticks ± SEM. Right: difference in mean percentage accuracy between no-TMR and TMR groups (B-E, G-H) or between two different types of task (F, I) shown using bootstrap-coupled estimation (DABEST) plots. Effect size for the difference between no-TMR and TMR groups was computed from 10,000 bias-corrected bootstrapped resamples: black dots, mean; black ticks, one-tailed 95% confidence interval; filled-curve, sampling-error distribution. (B) TMR had a significant effect on average memory performance across all tests (p = 0.007). (C-E) TMR had a significant effect on associative test performance (C, p < 0.001), with no effect observed for contextual tests (D, p = 0.696) or for the navigation test (E, p = 0.284). (F) The effect of TMR on associative tests was significantly greater than that observed for contextual and navigation tests (F, p = 0.007). (G, H) No significant effect of TMR was observed for directly trained tests (G, p = 0.426; one node step), while a significant effect of TMR was observed for all non-directly trained tests (H, p < 0.001; scene-scene, scene-cue, cue-cue, two-step). (I) The effect of TMR on the indirect two node step test was significantly greater than that observed on the direct one node step test (p = 0.010).

Fig. 4

Behavioural performance on each post-rest memory test included in the test phase. (A-H) Schematics and results for each of the 8 tests in the order that they were presented to participants, as shown in Fig. 3A: scene-scene, scene-cue, cue-cue, one node step, map1/map2, same/different, two node step inference, and shortest route. Left pair-plots: raw data points for no-TMR map (red; left) and TMR map (blue; right); each data point is mean accuracy for one participant; black dot, participant mean; black ticks ± SEM; black dotted line indicates chance level. Right: difference in mean percentage accuracy between no-TMR and TMR maps shown using bootstrap-coupled estimation (DABEST) plots. Effect size for the difference between no-TMR and TMR groups was computed from 10,000 bias-corrected bootstrapped resamples: black dots, mean; black ticks, one-tailed 95% confidence interval; filled-curve, sampling-error distribution. Purple: associative tests, Orange: contextual tests, Blue-Grey: navigation test. TMR was found to have a significant effect on scene-scene (A, p = 0.033), scene-cue (B, p = 0.007), cue-cue (C, p = 0.029), and two node step inference (G, p < 0.001), all of which fall under associative tests. No significant effect was found on one node step (D, p = 0.426), map1/map2 (E, p = 0.819), same/different (F, p = 0.469), or shortest route (H, p = 0.284). Notably, these statistical tests were all planned, as outlined in the pre-registration. (I) Regression analysis to control for potential confounding variables, as outlined in the pre-registration. Test performance accuracies for the two maps were used as the respective dependent variables, separately for the TMR and no-TMR conditions. Independent variables included the order of the map in training, age, gender, learning accuracy for that map, the colour of the map, and the associated auditory cue. The residuals were obtained for the two models, and the mean added back in, after which the difference in mean percentage accuracy between TMR and no-TMR groups (i.e. the TMR difference) was calculated and plotted. Top: each data point in the swarm plot is mean accuracy for TMR difference for one participant. Bottom: the TMR difference is shown using bootstrap-coupled estimation (DABEST) plots. Effect size for the TMR difference is computed from 10,000 bias-corrected bootstrapped resamples: black dots, mean; black ticks, one-tailed 95% confidence interval; filled-curve, sampling-error distribution. Consistent with the findings reported in Fig. 4A-H, when controlling for potential confounding variables, TMR was still found to have a significant effect on scene-scene (p = 0.001), scene-cue (p = 0.007), cue-cue (p = 0.022), and the two node step inference task (p < 0.001), all of which were categorised as associative tests. Similarly, no significant effect was found on one node step (p = 0.423), map1/map2 (p = 0.168), same/different (p = 0.471), or shortest route (p = 0.255).