Weak and strong memories are equally reactivated during counterfactual learning, but only weak memories are modified

Abstract

Adaptively updating memories about familiar others is an important social skill. Here, we examined whether strong or weak memories encoded in a social learning task are more prone to updating with new information when they are violated. Participants viewed well-known fictional characters’ emotional reactions to food items either repeatedly or only once. They then studied new events where some characters exhibited unexpected emotions to the same food items, violating initial memories. We found that when violated, strong existing memories, compared to weak memories, led to worse recall of the new, conflicting information, suggesting the resistance of strong memories to updating. Further, using a novel behavioral metric as a proxy for memory reactivation, we assessed the mechanistic role of reactivation in the updating of strong and weak memories. We showed that when encountering conflicting information, instead of reflecting their relative encoding strength, reactivation of weak memories reached a level similar to that of strong memories. Meanwhile, strong memories that were more reactivated showed higher likelihood to incorporate new information. Together, our findings provide insights into how existing social memories integrate with new information, a process shaped by memory strength and reactivation during new learning.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-65066-7.

Introduction

We are constantly learning about the people around us. However, it is challenging to maintain accurate knowledge about other people, as even those we know very well can change over time, and some changes may contradict our existing memories. For example, you might remember your friend frequently drinking a Chai latte, but recently found out that they no longer like this beverage. In such cases, beyond what you already knew about your friend, it is important to also keep their new beverage preference in mind. Successfully incorporating such new information into existing memories ensures that our behaviors during social interactions – e.g., choosing a drink for your friend at a café – are based on comprehensive and up-to-date knowledge. However, while some memories can be flexibly updated by integrating newly-encountered information, others may stay unchanged, and it remains an open question which factors modulate the malleability of an existing memory.

Extensive work in the field of cognitive neuroscience has suggested that memory updating is often triggered by new information that conflicts with our memory-based expectations, or in other words, mnemonic prediction errors^1–5. However, it is unclear precisely how the magnitude of a mnemonic prediction error is associated with the likelihood of incorporating new information in order to adjust existing memories. Regarding this open question, existing work has documented mixed findings. On the one hand, prediction violations induce surprise, and greater levels of surprise can result in increased incorporation of new information into existing knowledge^6,7. Prior work has also suggested that larger mnemonic prediction errors lead to better memory for the new information^8,9 and shift hippocampal processing towards new encoding to a greater extent¹⁰. Based on these studies, it is reasonable to hypothesize that violations of strong, as compared to weak, existing memories are more likely to result in memory updating with new information, given that violations of strong memories should induce larger prediction errors. On the other hand, and counter to this hypothesis, it has been revealed that existing memories that are strong and stable are less likely to undergo change^11–13. This is also broadly in line with prior work on memory interference, which highlights that existing knowledge can inhibit memories for newer experiences^14–20. Relatedly, it has been proposed that even in the presence of a prediction error signal, strong memories may downregulate the learning of new and contrasting information²¹. According to this view, strong existing memories may be preferentially retained over newly encoded information. Given these competing hypotheses, further investigation is needed to address the open question whether strong or weak prior memories are more vulnerable to updating by incorporating new information when violated.

Another factor that plays an important role in the process of memory updating is the reactivation of existing memories. A large body of literature has suggested that a memory becomes temporarily labile when it is reactivated but can otherwise stay relatively stable^22–26. Therefore, reactivation creates an opportunity for pre-existing memories to incorporate new information and potentially serves as a prerequisite for memory change^2,4,5,27,28. Some recent work has taken a step further to examine how varying levels of reactivation might induce qualitatively different changes to the reactivated memory^29–32. However, what determines the degree of reactivation of a memory in the first place, especially at the time when the memory is violated, is understudied. Some work has suggested that learning new information that is congruent, rather than incongruent, with existing memories is associated with enhanced reactivation of these memories³³. In contrast, other findings have shown that when memory-based predictions are violated, strong reactivation of existing memories is needed to induce a mismatch signal in the hippocampus³⁴. Further, while it is reasonable to hypothesize that strongly encoded memories should be overall more reactivated than weakly encoded memories^35,36, whether and how encoding strength interacts with mnemonic prediction errors in modulating memory reactivation remains largely unknown. Additionally, level of memory reactivation has been predominantly assessed with neuroimaging and electrophysiology approaches, yet reactivation supports and is closely associated with retrieval behaviors^36–45. To build on this work, developing behavioral measures that tap into memory reactivation processes can offer valuable insights but remains a challenge for memory researchers.

Critically, these open questions are significant in many domains including social learning, as memories that need updating often pertains to information about other people. Decades of research has investigated how we remember information that is congruent versus incongruent with existing person knowledge^3,46–51. Mixed findings have been documented, with some studies suggesting better memory for expectation-congruent information than expectation-incongruent information (e.g.,^52,53) and others showing the opposite (e.g.,^46,54). While it has also been a longstanding idea that the strength of existing person knowledge plays a role in shaping the memory for the new information when expectations are violated^3,48,49, prior work has relied on inference about the knowledge strength based on social factors such as interpersonal distance (e.g., friend vs. stranger;⁵⁵) and group membership (ingroup vs. outgroup stereotypes;⁵⁶). Only limited empirical work to our knowledge has directly manipulated the encoding strength of person knowledge in an expectation violation task where social motivation is minimal⁵⁷. Moreover, while reactivation has been considered a key mechanism associated with memory lability^22–26, how it modulates the updating of person knowledge in a social learning context and how this effect might differ for strong versus weak existing knowledge remain to be examined.

The current study aims to bridge these gaps. Specifically, we examined whether, upon encountering mnemonic prediction errors, strong and weak memories obtained in a social learning context are differentially updated by successfully incorporating new, contrasting information. We also explored the mechanistic role of memory reactivation in this process. Participants learned food preferences of fictional characters by viewing each character’s dynamic emotional reactions to a food item. Importantly, information about some characters was studied three times while others only once, leading to memories of different levels of strength. We then violated half of these strongly and weakly encoded memories by presenting participants with new events where the studied characters showed unexpected (i.e., opposite) emotional reactions to the same food items. We later tested participants’ memory for these new, contrasting emotional information. Moreover, we developed a behavioral metric to examine reactivation upon memory violation in our experiments. Specifically, we inferred levels of reactivation based on the continuous precision of participants’ memory for an episodic detail from the prior related memory, retrieved immediately after memory violation. We then examined how this retrieval precision, as a proxy for reactivation magnitude, reflected prior memory strength and related to new learning.

We find that memories about other people that were repeatedly encoded are more resistant to incorporating conflicting information compared to those encoded only once, such that the new information is less retained when it contradicts strong versus weak memories. Interestingly, we found that weak memories are reactivated to the same extent as strong memories when violated. However, reactivation of strong prior memories, in particular, tracks their likelihood of integrating new, contrasting information on a trial-by-trial basis.

Methods

Participants

Data for the current manuscript are from three rounds of collection involving highly overlapping experimental conditions, with minor differences noted in the Supplementary Methods. A total of 206 online participants were recruited on Prolific across the three experiments, separately conducted in April 2021 (Experiment 1: N = 54), September 2021 (Experiment 2: N = 75), and May 2022 (Experiment 3: N = 77). All participants reported normal or corrected-to-normal vision, fluency in English, residency within the United States, no history of a primary psychiatric disorder, and not taking neurological medications. Twenty-six participants were excluded from all analyses due to failure to perform the experimental tasks as instructed, including skipping a substantial number of trials (N = 7) and exhibiting extremely poor learning (N = 19). Data from the remaining 180 subjects were reported in Results (mean age = 27.01 years, SD = 4.70 years, range: 18–35 years; 83 females, 96 males, and 1 reported non-binary gender; see Supplementary Methods for a detailed breakdown of participants demographics, exclusions, and sample size rationale for each experiment). Within the remaining sample, we additionally excluded 2 participants from analyses of the food-character associative memory test due to unreasonably fast response times (median < 0.35 s; both participants performed below chance-level on this test). All participants provided informed consent prior to participation, and received monetary compensation on study completion. The study protocol was approved by the Columbia University Institutional Review Board, and all experiments were performed in accordance with relevant guidelines and regulations.

Materials

Stimuli portraying well known characters

Dynamic GIFs (Graphics Interchange Format images; i.e., short, looped video clips) and static images portraying 36 famous characters, including 18 animated characters (e.g., Pikachu) and 18 real persons (e.g., Sherlock Holmes), were collected from the Internet and used in the experiments. We selected both animated and real-person characters to represent a wide range of character types across common media formats.

GIFs

In all three experiments, two different GIFs featuring each famous character were presented during encoding, one for each encoding phase (see Procedures). In each GIF, the featured character exhibited either a positive or a negative emotion through their facial expressions (e.g., smile) and body movements (e.g., dancing). Across the two encoding phases, the two GIFs of a character either showed emotions of the same valence (i.e., both positive or both negative) or opposite valence (i.e., one positive and one negative), forming the “same” and “change” emotion conditions, respectively (Fig. 1; see also Procedures).

Fig. 1.

Experimental design. Participants completed an encoding section and a testing section. Encoding section consisted of two phases: initial learning (Phase 1) and memory violation (Phase 2). During Phase 1, half the trials were presented three times, forming the “prior strong” cognition, and the other half only once, forming the “prior weak” condition. Each trial during Phase 1 consisted of three steps: shape encoding, shape working memory test, and associative encoding. During Phase 2, all trials were presented once only. Each Phase 2 trial consisted of two steps: new associative encoding and reactivation-based shape test. From Phase 1 to Phase 2, half the studied characters changed their emotion exhibited in the GIFs (video clips) to the same assigned food items (“change” condition), while the other half showed the same emotions (“same” condition). Note that all GIFs in Phase 2 were different from Phase 1 in both “change” and “same” conditions. After a short delay following encoding, participants then performed a two-alternative forced-choice food-character associative memory test, an emotion memory recall test, and a specific video/GIF recognition test.

In Experiments 2 and 3, two additional GIFs of each character were used during the testing section. Specifically, for each GIF presented during encoding, a novel GIF featuring the same character showing emotion of the same valence served as the test lure during the specific video recognition test (see Procedures for details; this test was not conducted in Experiment 1).

For each GIF stimulus used in the study, the valence of the exhibited emotion was validated by an independent sample of participants in a separate norming experiment (see Supplementary Methods for details).

Static images

One colored static image of each character was used during the testing section of all experiments, with the background of the image removed.

Images of food items

A total of 36 colored photographs of common food items, including 18 fruits and 18 vegetables, were used in the experiments and were collected from the Internet, with backgrounds removed.

Picnic blankets and shape stimuli

Thirty-six images of picnic blankets were created and used during the encoding section of all experiments. Each blanket had a white background and a black border, together with fourteen identical black shapes aligned inside the border (Fig. 1). Importantly, the shapes presented on each blanket were different from the ones on any other blanket in the stimulus set. Therefore, a total of 36 different shapes, one type presented on each blanket, were used in the experiments. All 36 shapes were selected from the Validated Circular Shape space⁵⁸, sampled at intervals of 10 degrees distance; the degree distance between any two different shapes represents the level of perceptual dissimilarity between them.

We also constructed a shape wheel including all 36 shapes, which was presented during the shape tests during Encoding Phases 1 and 2 (Fig. 1; see also Fig. 4A). On the shape wheel, the shapes were arranged based on their perpetual (dis)similarity, such that any two neighboring shapes were 10-degree different from each other. Therefore, shapes closer to each other on the wheel were perceptually more similar than shapes further apart.

Fig. 4.

Performance on the reactivation-based shape test during Phase 2 of encoding. (A) We examined the magnitude of the errors in the responses on the reactivation-based shape test by computing the degree differences between the responses and the target shapes. A response that showed an offset greater than 90 degrees from the target shape was considered as minimal memory for the correct shape. (B) Distribution of shape selection offset across all conditions. The dotted line in the middle of each sub-panel represents the target shape. The magnitude of memory deviation increases as the response moves further away from the target shape. Responses with an error greater than 90 degrees were considered as minimal memory and excluded from the computation of reactivation-based memory precision. (C) Reactivation precision across conditions. We computed the average deviation for responses on the Phase 2 shape test that showed an offset within 90 degrees away from the target shape. Lower values indicate more precise reactivation. Thick short bars represent mean values. Error bars show within-subject standard errors. Ⓧ significant interaction; ns: not significant; **p < 0.01.

Emotion slider scales

Continuous emotion slider scales were used in both encoding and testing sections in all experiments for participants to rate or to recall emotional information (see Procedures). Each scale ranged from “extremely negative” on the leftmost end (value coded as 0) to “extremely positive” on the rightmost end (value coded as 100), with the slider positioned initially at the midpoint denoting a neutral emotion (value coded as 50). The numerical value of the slider position was only used during data analyses but never shown to the participants.

Procedures

In all experiments, participants first completed two encoding phases: initial learning and memory violation. They then performed a 5-minute distractor task (odd/even number judgement task). They then completed the testing section following the distractor task (Fig. 1). The procedures of Experiments 2 and 3 were entirely identical, and both contained minor differences from that of Experiment 1. Below, we focused on reporting the procedures shared across all experiments, and we noted the deviations in Experiment 1 paradigm where applicable.

Encoding

The encoding phase adopted a 2 × 2 within-subjects design crossing Phase 1 encoding strength conditions (prior weak, prior strong) and emotion conditions (change, same) to yield a total of four conditions.

Phase 1: initial learning

During initial learning, participants viewed different food-GIF associations, each presented atop an image of a unique picnic blanket. To manipulate the encoding strength for these associations, half the Phase 1 trials were repeated three times, which comprised the prior strong condition; the other half were presented only once, which formed the prior weak condition. Trials of both conditions were presented in an interleaved manner. Participants in all three experiments studied 36 unique associations during Phase 1. Experiment 1 included 12 different associations in each of prior strong and prior weak conditions (24 associations and 48 trials in total: 12 × 3 + 12 × 1; another 12 associations in Experiment 1 were used in a condition irrelevant to the current study, see Supplemental Methods); Experiments 2 and 3 included 18 associations in each of prior strong and prior weak conditions (36 associations and 72 trials in total: 18 × 3 + 18 × 1).

Prior to Phase 1, participants were told the cover story that famous characters featured in GIFs were going on picnics. Each character would bring their customized picnic blanket and was assigned with a food item that they either liked (i.e., showed a positive emotion in the GIF) or disliked (i.e., showed a negative emotion). Participants were instructed to memorize all aspects of each picnic event and that their memory would be tested later.

Each Phase 1 trial unfolded in 3 steps: (a) shape encoding, (b) shape working memory test, and (c) associative encoding (Fig. 1; all steps repeated in the prior strong condition). During shape encoding, a picnic blanket featuring one specific shape was presented alone for 1.5 s before it disappeared from the screen. Immediately after, participants performed a shape working memory test, in which they were presented with the shape wheel and asked to select (click on) the shape that was on the blanket they had just seen (Figs. 1 and 4A). Participants were allowed to make a maximum of two attempts, such that the screen automatically advanced after receiving two mouse clicks, with only the second response being considered. Alternatively, participants could also submit their response by clicking on the “continue” button when they needed no more than one attempt. This test was self-paced. Following the shape test, the picnic blanket re-appeared, together with an image of a food item and a GIF featuring a famous character. All three elements—the picnic blanket, the food item, and the GIF—were presented together for 8.5 s, during which participants were asked to rate the emotion of the character in the GIF, using an emotion slider scale presented at the bottom of the screen. Phase 1 trials were designed this way to ensure attention was oriented to all aspects of the encoding materials (picnic blanket shape, gif character identity and emotion, and the food item), as all information was important for both the memory violation phase (Phase 2) and the memory tests.

Phase 2: memory violation

After all Phase 1 trials, participants studied another set of picnic events featuring the same characters they had seen before. Importantly, each character was assigned with the same food item as in Phase 1, but their emotion towards the food either stayed consistent with that during Phase 1 (“same” condition; e.g., emotion stayed positive across Phases 1 and 2), or showed an opposite valence (“change” condition; e.g., emotion changed from positive in Phase 1 to negative in Phase 2; Fig. 1). Half the characters were assigned to the “same” condition while the other half to the “change” condition. All Phase 2 trials were presented once only (24 trials in Experiment 1, 36 trials in Experiments 2 and 3).

Each Phase 2 trial consisted of 2 steps: (a) new associative encoding, and (b) reactivation-based shape test (Fig. 1). During new associative encoding, participants were presented with a novel GIF of a familiar character paired with the same food item that was previously associated with the character during Phase 1; the picnic blanket was no longer presented. The food-GIF pair was presented for 6 s, during which participants were asked to rate the character’s emotion exhibited in the new GIF using the emotion slider scale. It is worth emphasizing that the GIFs presented in Phase 2 were always different from those in Phase 1, regardless of the emotion conditions. While the participants were instructed that they would view novel GIFs during Phase 2, they were unaware of the emotion manipulations. Following new associative encoding, the food-GIF pair continued to be presented, and a shape wheel appeared on screen. Here, participants were asked to recall the shape that was on the given character’s picnic blanket that they had seen during Phase 1 (not presented during Phase 2), and to select the recalled shape on the wheel. We reason that memory retrieval of this episodic detail engages the pattern completion mechanisms that are supported by the underlying reactivation of the related experiences^36,39,59,60. Similar to the Phase 1 shape working memory test, participants can make a maximum of two self-paced attempts or proceed after one attempt. Given the overlap between the stimuli in Phases 1 and 2, we probed memory for the no-longer-presented shape information (rather than other elements from the Phase 1 events) as a means to assess reactivation during new learning to reduce potential interference between the concurrently presented stimuli and the tested content.

Across both encoding phases, the pairing between GIFs, food items, and picnic blankets (shapes), as well as the assignment of the character-food-blanket associations to the prior memory strength conditions (prior weak vs. prior strong) and emotion conditions (same vs. change), were counterbalanced across participants.

Testing

Food-character association

Following the distractor task, participants were first tested on their memory for the food-character associations in a two-alternative forced-choice test. On each trial, participants were presented with a studied food item and asked to select the character that was paired with the food from two options (static images of two characters; the foil was a familiar character from the same encoding condition as the target). Each trial had a response window of 4 s, with a 1-second intertrial interval (ITI).

Initial & updated emotions

Participants then completed the emotion recall test. On each trial, participants were presented with a food-character (static image) pair; they used separate slider scales to recall the initial (Phase 1) and new (Phase 2) emotions exhibited by the character towards the food item. This test was self-paced with a maximum response window of 10 s on each trial.

Specific video (GIF) recognition

Participants then performed a specific video (GIF) recognition test. They were presented with pairs of GIFs of the same character showing emotion of the same valence (e.g., two different GIFs of a happy Pikachu). From each pair, they were asked to choose the GIF that had been shown during encoding. Each trial had a response window of 4 s with a 1-second ITI. All studied GIFs from both encoding phases were tested in a randomized order.

For each studied character, the assignment of a GIF to be presented during encoding or to serve as a test lure was counterbalanced across participants. This test was conducted in Experiments 2 and 3, but not in Experiment 1.

Analysis

Data synthesis across experiments

All analyses reported in the main text were performed on data collapsed across all three experiments due to their highly overlapping experimental procedures. This allowed us to obtain a large sample and to gain increased statistical power (but see also Supplemental Data for results from each individual experiment). However, it is worth noting that we planned sample sizes separately prior to conducting each individual experiment. The aggregated data, which was larger than any individual experiment, reflected multiple rounds of collection, rather than a one single experiment. Taking into consideration that the experiments were conducted serially and that there may be potential differences in results between experiments, we have adopted a meta-analytic approach and examined whether and how experiment as a factor (1 vs. 2 vs. 3) interacted with experimental conditions in affecting performance on each memory test using model comparisons. To this end, we fitted subject-level performance on each memory test in a pair of two mixed-effects models (see also “Memory performance across conditions” below): (i) a model including experiment and its interaction with the other fixed-effects regressors as predictors; and (ii) a model in which we controlled for experiment but did not include its interaction with the other regressors. We first compared how well each dependent variable (specified in Results) was predicted by these two models using a likelihood-ratio test, and then reported the results from the better-fitting model. When Model i both significantly outperformed Model ii and revealed any significant interactive effect of experiment, we further unpacked the details of such interaction(s). When the two models did not differ significantly in their predictive power, we reported results from Model ii to avoid the cost of extra degrees of freedom. From the winning model (or Model ii when no difference was found between models), we computed the average effect of any fixed-effects regressor across multiple levels of other regressors (e.g., combined effects across experiments) with type III ANOVAs using Satterthwaite’s method⁶¹. We also performed comparisons of marginal means between conditions of interest using the “emmeans” package in R^62,63. For completeness, we separately conducted all analyses within each individual experiment and reported these results in Supplemental Data (see Supplemental Materials). We also encourage readers to evaluate the reported effects both in the aggregated sample across all experiments and within individual experiments, as some of the results reached statistical significance in the larger, combined sample, but did not replicate across all individual experiments.

Shape and emotion learning performance

We first confirmed that during encoding, participants were able to correctly identify (1) the target shapes when performing the shape working memory test during Phase 1, and (2) the emotional valence exhibited in the GIFs during both encoding phases. To determine the perceived valence from participants’ emotion ratings, we binarized the slider values into positive (any values greater than 50) and negative (any values smaller than 50) emotions. We then compared the rated valence with the validated valence based on the norming data of each corresponding GIF (see Supplementary Methods for details about the norming experiment) to determine the valence perception accuracy.

Memory performance across conditions

We used mixed-effects models to compare the subject-level mean performance across experimental conditions, separately for the reactivation-based shape test during Phase 2 of encoding and all three memory tests during the testing section. In these models, we included Phase 1 encoding strength condition (prior weak vs. prior strong), emotion condition (change vs. same), and their interaction as fixed-effects regressors, and random intercepts for each participant. We further assessed whether these effects varied across experiments using a model comparison approach described above (see “Data synthesis across experiments”).

For performance on the reactivation-based shape test, our primary measure of interest was memory precision of the shape information, which was operationalized as the absolute degree difference between the target and the response. We used this measure to index the continuous level of memory reactivation as reactivation strength has been shown to predict retrieval accuracy and vividness^64–66. Given that the shape test was highly difficult, we limited our analyses to trials where participants (1) had properly learned the corresponding shape information during Encoding Phase 1 (0-12.5% of trials excluded across participants; 4.08% on average), and (2) demonstrated potential memory signals of that shape during Encoding Phase 2 (4.17%-63.89% of trials excluded across participants; 42.75% on average). Specifically, a trial during Phase 2 shape test would only fulfill these requirements when the shape responses on both the corresponding Phase 1 (one trial in the prior weak condition, or all three repetitions in the prior strong condition) and Phase 2 test trials were located within the same half-circle on the shape wheel as the target. Although set arbitrarily, the half-circle threshold corresponds to a chance level of 0.5 in selecting a shape on the correct side under random guessing (i.e., | target – response | ≤ 90 degrees; Fig. 4A). If a response was above 90 degrees apart from the target shape during either Encoding Phase 1 or Phase 2 shape test, we considered the corresponding shape information to be poorly learned or minimally remembered (i.e., forgotten), respectively; and we excluded that Phase 2 trial from the computation of reactivation-based memory precision (see Fig.S1 & Fig. 4B). Rather than introducing substantial noise into the reactivation precision measure by including all trials for this analysis, we applied these exclusions in order to ensure that only trials where memory reactivation was sufficiently expressed in the associated behavior were considered.

In assessing performance on the emotion recall test, we only included trials where participants had correctly identified the given characters’ emotional valence on all relevant encoding trials during both Encoding Phase 1 and Phase 2. The exclusion criteria for the computation of reactivation-based memory precision and emotion recall performance were determined a priori.

Relating shape reactivation precision with memory for changed emotional valence

Within the emotion “change” condition, we performed a trial-level mixed-effects model to predict whether the recall success of the updated emotional valence (Phase 2 emotion) during the emotion recall test (remembered = 1, forgotten = 0), with the reactivation-based memory precision for the related prior event as measured with Phase 2 shape test. The model used a logistic linking function. Reactivation-based shape memory precision (numeric variable; mean-centered within participants), Phase 1 encoding strength (two-level categorical variable: prior weak, prior strong; dummy-coded), and their interaction were included as fixed-effects regressors, and random intercepts were included for each participant. To avoid overfitting due to limited number of trials per sampling unit, we did not include complex random slopes per participant, nor did we include experiment as an additional regressor (but see Supplemental Data for separate model results in each individual experiment).

Results

Shape and emotion learning performance

During Encoding Phase 1, participants were able to distinguish the picnic blanket shape patterns and selected a shape that fell within the same half-circle as the target on 98% of the trials on the shape working memory test (vs. chance level 0.5: t(179) = 42.19, p < 0.001, 95%CI[0.97, 0.98], Cohen’s d = 14.33; Fig. S1). Participants also successfully recognized the emotional valence exhibited in the GIFs, with an accuracy of 0.95 rating the correct valence (positive or negative) during Encoding Phase 1, and an accuracy of 0.94 during Encoding Phase 2; both were significantly above chance (0.5; Phase 1: t(179) = 89.21, p < 0.001, 95%CI[0.94, 0.96], Cohen’s d = 6.65; Phase 2: t(179) = 73.96, p < 0.001, 95%CI[0.93, 0.95], Cohen’s d = 5.51).

Repetition during initial learning strengthens memory

Participants studied half the Phase 1 trials three times, viewing all elements repeatedly, while the other half only once. Here, we examined whether encoding repetition strengthened subsequent item and associative memory for Phase 1 events. We separately assessed participants’ memory accuracy on (1) specific video recognition test for GIFs viewed during Phase 1, (2) food-character associative memory test (note that all food-character pairs were presented once again during Encoding Phase 2), and (3) emotion recall test for the studied characters’ initial emotional valence towards the assigned food items. As expected, participants showed significantly better memory for the repeatedly studied events than those studied only once (Phase 1 video recognition: F(1, 414) = 148.98, p < 0.001, η² = 0.265; character-food associations: F(1, 531) = 15.50, p < 0.001, η² = 0.028; Phase 1 emotional valence recall: F(1, 531) = 111.86, p < 0.001, η² = 0.174; Fig. 2; see Table S1 for complete model outputs and model comparisons). These results confirmed that our repetition manipulation during initial learning led to robust differences in memory strength. The effect of repetition on memory for Phase 1 events did not significantly vary across experiments (Phase 1 encoding strength × experiment: all p > 0.07).

Fig. 2.

Repetition during initial learning strengthened memory. Participants showed significantly better memory for the specific GIFs presented during Phase 1 (left), food-character associations (middle), and initial emotional valence of the characters exhibited towards the food items (right) in the prior strong (thrice-encoded during Phase 1) as compared to prior weak (once-encoded during Phase 1) condition. Each dot represents an individual subject. Dotted lines represent chance levels. Thick short bars represent mean values. Error bars show within-subject standard errors. *p < 0.01, **p < 0.01, ***p < 0.001.

Emotion consistency across encoding benefits subsequent recall of Phase 1 emotions

We found that recall accuracy of the studied characters’ initial emotion was also affected by whether the emotion changed or stayed the same from Encoding Phase 1 to Phase 2 (F(1, 531) = 15.22, p < 0.001, η² = 0.028), in both prior weak (b = 0.042, p = 0.019, 95%CI[0.007, 0.077]) and prior strong (b = 0.056, p = 0.002, 95%CI[0.021, 0.091]) conditions (Fig. 2). Specifically, when a character’s emotional valence stayed consistent across Phases 1 and 2, participants showed a significantly higher accuracy recalling the character’s initial (Phase 1) emotion, as compared to.

when emotion changed. However, this effect varied across experiments, such that including the interactive effects of experiment significantly improved model fit in predicting initial emotion recall (χ²(6) = 13.62, p = 0.034), driven by a significant interaction between experiment and emotion condition (F(2, 531) = 3.85, p = 0.022, η² = 0.014). Specifically, emotion consistency significantly improved the recall of initial valence in Experiment 2 (b = 0.082, p < 0.001, 95%CI[0.043, 0.121]) and Experiment 3 (b = 0.069, p < 0.001, 95%CI[0.031, 0.107]), but not in Experiment 1 (b = 0.004, p = 0.87, 95%CI[-0.054, 0.046]).

Combined across Phase 1 encoding strength conditions, a similar memory benefit of emotion consistency (same > change) was found for the recognition accuracy of the specific Phase 1 GIFs (F(1, 414) = 4.05, p = 0.045, η² = 0.010). However, this effect was not significant when examined within either prior weak (although a trend was noted: b = 0.023, p = 0.080, 95%CI[-0.002, 0.049]) or prior strong condition (b = 0.014, p = 0.28, 95%CI[-0.012, 0.040]) alone. Memory for the food-character associations did not differ across emotion conditions (same vs. change; F(1, 531) = 0.052, p = 0.82, η² = 0.0001). Adding experiment as an interaction term did not improve the explanatory power of the models predicting Phase 1 video recognition (χ²(3) = 6.26, p = 0.099) or food-character associative memory (χ²(6) = 4.00, p = 0.68).

Strong memories are resistant to updating when they are violated

After confirming that repetition effectively strengthened memory for different aspects of Phase 1 events, we next asked our central question: Are strong or weak memories more prone to be updated when encountering new, contrasting information? Of particular interest was the updating of memory for the studied characters’ emotions towards the assigned food items. To test this, we examined participants’ subsequent recall performance for the new emotion a character exhibited during Encoding Phase 2. We reasoned that accurate recall of Phase 2 emotion that violated existing memory would indicate successful incorporation of this new information in updating the related prior knowledge. (For these analyses, we focused on reporting results within the “change” condition in the main text; for performance across all experimental conditions, see Supplemental Data and Fig. S2; see also Table S1d & S1e for model outputs).

We assessed Phase 2 emotion recall performance with two measures. First, we examined participants’ accuracy in recalling the binary emotional valence when the related prior events were encoded three times (prior strong) versus only once (prior weak). We found that when the emotion changed from Phase 1 to Phase 2, participants were significantly more accurate at recalling the new emotional valence in the prior weak condition than prior strong condition (b = 0.046, p = 0.026, 95%CI[0.006, 0.087]; Fig. 3 middle). Second, we used the continuous emotion scale to measure the precise memory for these new emotions. Specifically, this was measured with the continuous difference score between the emotion rating a participant provided during encoding and the corresponding emotion rating recalled at test. Consistent with results for the binary valence, participants showed less precise memory (larger encoding-recall difference scores) of the new emotions when the corresponding prior association was strongly, versus weakly, encoded (b = 2.47, p = 0.040, 95%CI[0.109, 4.82]; Fig. 3 right). Therefore, as compared to prior memories formed via one-shot learning, repeatedly-learned memories might be more resistant to incorporating new, conflicting information, which resulted in worse episodic memory for both coarse- and fine- grained information encountered at the new events. Interactive effects of experiment did not significantly affect model fit predicting new emotion recall performance (both p > 0.064).

Fig. 3.

Memory of Phase 2 emotion in change condition. When the emotion of a character changed from Encoding Phase 1 to Phase 2, participants were significantly better at recalling the new emotion – both for the binary valence (middle) and for the precise emotion ratings (right) – when their prior memories about the characters were weakly encoded than when they were strongly encoded. Dotted line in the middle plot represents chance levels. Thick short bars represent mean values. Error bars show within-subject standard errors. *p < 0.05.

We also examined whether strong prior memories led to poorer recall of the emotional information specifically – the to-be-updated element in the existing memories, or more generally of all new information presented during Phase 2. To this end, we examined whether Phase 1 encoding strength also affected subsequent item memory for the specific GIFs viewed during Phase 2. We did not find a significant difference in recognition accuracy for Phase 2 GIFs between prior strong and prior weak conditions when emotion changed (b=-0.024, p = 0.13, 95%CI[-0.055, 0.007]; Figure S3; see Supplemental Data for results across all conditions; see Table S1g for model output). Note that the null effect here could be partially due to reduced power, as fewer participants performed the GIF recognition test (N = 139; Experiments 2 & 3) than the emotion recall test (N = 180; all three experiments). But it is also possible that instead of interfering with new learning across the board, strong existing knowledge leads to reduced sensitivity to the incompatible information in particular in the upcoming events.

Memory violation overrides the effect of encoding strength on reactivation levels

Having showed that prior memories that were weakly versus strongly encoded differed in their propensity to incorporate new, contrasting information, we next examined whether the extent of such memory updating may be related to reactivation of the prior learned associations. To test this, we developed a behavioral assay of memory reactivation. Specifically, we used memory precision for the encoded shape information assessed during each violation trial of encoding Phase 2 as a continuous behavioral metric of reactivation.

The reactivation-based shape memory test was a highly difficult task, but participants performed above chance (although with small effect sizes) in selecting a shape that fell within the same half-circle as the target (i.e., deviation below 90 degrees; Fig. 4A), both when the shape information was encoded three times (prior strong: t(179) = 6.05, p < 0.001, 95%CI[0.55, 0.59], Cohen’s d = 0.45) and only once (prior weak: t(179) = 4.14, p < 0.001, 95%CI[0.52, 0.56], Cohen’s d = 0.31). Here, we first asked whether and how our primary dependent variable of interest on this test – reactivation-based memory precision – might be affected by the experimental manipulations. Results revealed significantly higher reactivation precision (lower degree deviation between responses and target shapes) when Phase 1 events were strongly versus weakly encoded (F(1, 530) = 4.34, p = 0.038, η² = 0.008; see also Table S1a for model output), but we also found a significant interaction between Phase 1 encoding strength and emotion conditions (F(1, 530) = 6.6959, p = 0.0099, η² = 0.012; Fig. 4B and C). Specifically, repetition improved reactivation-based memory precision when emotion stayed the same (b=-5.23, p = 0.001, 95%CI[-8.34, -2.12]), which was consistent with the general strengthening effect of encoding repetition as aforementioned. But surprisingly, when the emotion changed from Phase 1 to Phase 2, memory for the shape information was not significantly more precise in the prior strong versus prior weak condition (b = 0.566, p = 0.72, 95%CI[-2.55, 3.68]), despite the repeated learning of the picnic blankets in the prior strong condition. Additional exploratory comparisons of reactivation precision between emotion conditions revealed opposite trends when prior memory was strongly (marginally more precise in same vs. change: b = 2.78, p = 0.080, 95%CI[-0.33, 5.88]) and weakly encoded (marginally less precise in same vs. change: b=-3.02, p = 0.058, 95%CI[-6.13, 0.097]), although the effects did not reach statistical significance.

Notably, these results differed across experiments (experiment × Phase 1 encoding strength: F(2, 530) = 3.04, p = 0.049, η² = 0.011; experiment × Phase 1 encoding strength × emotion: F(2, 530) = 3.19, p = 0.042, η² = 0.012), and including the interactions between experiment and other regressors significantly improved model fit (χ²(6) = 18.29, p = 0.006; see Table S1a for complete model results). Specifically, in each of Experiments 1 and 2, Phase 1 encoding repetition significantly improved shape reactivation precision only in the emotion “same” condition (Experiment 1: b=-9.8, p = 0.002, 95%CI[-16.1, -3.5]; Experiment 2: b=-7.14, p = 0.005, 95%CI[-12.1, -2.21]), but not in the emotion “change” condition (Experiment 1: b = 3.99, p = 0.218, 95%CI[-2.37, 10.3]; Experiment 2: b=-2.96, p = 0.240, 95%CI[-7.89, 1.98]), similar to the patterns observed in all data collapsed across experiments. In Experiment 3, however, reactivation precision did not differ between prior strong and prior weak conditions in either emotion “change” (b = 0.67, p = 0.78, 95%CI[-4.09, 5.43]) or “same” (b = 1.26, p = 0.60, 95%CI[-3.5, 6.02]) condition.

Updating of strong memories is modulated by reactivation upon violation

We next examined whether successful updating of existing memories – as indicated by subsequent memory for the new, contrasting information – can be predicted by the level of memory reactivation of related events upon violation. To do this, in both prior strong and prior weak conditions, we assessed the trial-specific relationship between the shape reactivation precision measured during a Phase 2 picnic event and the subsequent recall success of the updated emotional valence learned from that given new event. Results revealed that upon memory violation, when a Phase 1 event was repeatedly encoded (prior strong condition), more precise retrieval – interpreted as stronger reactivation – of that prior related event predicted a higher likelihood of remembering new emotional information contrasting with existing memory (b=-0.007, p = 0.040, 95%CI[-0.014, -0.0003]; Fig. 5; see Table S2 for complete model results). This result might suggest that upon encountering mnemonic prediction errors, the likelihood of updating a strong memory can be predicted by the continuous level of memory reactivation. This relationship was not found in the prior weak condition (b=-0.001, p = 0.75, 95%CI[-0.007, 0.005]). However, given that the interaction between Phase 1 encoding strength and reactivation precision did not reach statistical significance (b = 0.006, p = 0.200, 95%CI[-0.003, 0.015]), the difference between prior strong and prior weak conditions should be interpreted with caution.

Fig. 5.

Predicting memory for new emotion with reactivation precision upon violation. We conducted a trial-level mixed-effects model to examine in the emotion “change” condition, whether reactivation precision measured from Phase 2 shape test predicted the probability recalling the correct valence of the new, contrasting emotion, when prior memories were weakly and strongly encoded. Ribbons represent 95% confidence intervals. Each dot (aligned at 1 or 0 on the y-axis) represents an individual trial. Because the dependent variable is binary (i.e., valence memory outcome: remembered = 1; forgotten = 0), all dots fall at either 1 or 0. *p < 0.05.

Discussion

In the current study, we examined when and how memories of varying encoding strengths are updated in a social learning context through the successful integration of newly-encountered, contrasting information. Significant differences are revealed between strong and weak memories in their susceptibility to incorporating new information when encountering mnemonic prediction errors. We show that incoming information conflicting with strong prior memories will be less remembered, as compared to information that violates weak, or newly formed, memories. However, despite strong memories’ general tendency to resist change, the updating of a specific strong memory is correlated with the strength of memory reactivation at the time of new learning. Interestingly, reactivation of a weak memory during new learning is not directly linked to its updating. However, upon violation, overall levels of reactivation of weak memories – in spite of their low encoding strength – are surprisingly comparable to that of strong memories.

Together, our findings underscore that memory updating is an adaptive process contingent on the plasticity of existing memories, which is modulated by encoding strength and potentially gated by reactivation upon memory violation.

Building upon prior work demonstrating that mnemonic prediction errors can trigger new learning^8,10,67–70, our study highlights that this effect depends on the strength of the memories being violated. Specifically, new information that violates strongly encoded memories is remembered less accurately at both coarse- and fine-grained levels, as compared to information contrasting with the weaker memories. Critically, this finding is inconsistent with the notion that strong prediction errors or higher levels of surprise lead to more updating of existing memories^6–9,71. In our experiments, violations of the strong memories (i.e., prior strong condition) should have led to larger prediction errors than violations of weak memories (i.e., prior weak condition), but participants showed better memory for the new information in the prior weak condition. Therefore, instead of positively reflecting the magnitude of prediction errors or surprise, the level of new learning in the current study may be linked to the flexibility of prior knowledge. Specifically, repeated events may be represented in memory as stable knowledge that resists change, while the once-encoded memories show less rigidity and, hence, greater flexibility. Supporting this idea, our prior work has suggested that encoding repetition increases post-encoding cortical reactivation, providing evidence for rapidly established and stabilized cortical representations⁷². It is possible that as compared to these cortical memory representations, memories for the once-encoded, novel associations, which are still highly hippocampal-dependent, are also less stable and thus more sensitive to conflicting information and more likely to update.

In the current experiments, the task goal was to acquire and potentially update knowledge about fictional characters’ food preferences. Although the stimuli in the current study were not participants’ socially significant individuals from their real lives (e.g., family or friends), our task nonetheless created a social learning context. When learning about others, people’s behaviors can be influenced by the prioritization of pursuing accuracy versus efficiency^48,73. Here, we propose that the weight of the two can be affected by the strength of existing knowledge pertaining to the target individuals^{3,49,57,74,75}. Specifically, for individuals with whom we have limited experiences, it is important to continuously gather information at every possible learning opportunity, until we gain an accurate and comprehensive understanding of their actions. Actively integrating new information into existing memories can help eliminate uncertainty^76,77 and ultimately better guide future social decisions⁷⁸. In contrast, when interacting with those whom we have repeatedly encountered, we may make efficient judgements based on existing memories instead of constantly updating our knowledge⁷⁵. This is because (1) recurring past experiences have allowed us to build confidence about the pre-existing knowledge, and (2) frequent future interactions are expected, making it costly to keep recalibrating this knowledge⁷⁹. Therefore, maintaining stable representations resistant to change may enable rapid selections of social actions during potential future interactions. Supporting this view, past work has shown that unexpected new information conflicting with strongly encoded social knowledge may be irrationally discounted due to confirmation bias^80–82, or separately accommodated by generating auxiliary explanations (i.e., in line with Bayesian inference;^21,74,83). An example of the latter could be that after repeatedly learning Pikachu loves blueberries, one might justify Pikachu’s sudden frustration with blueberries by proposing that Pikachu may be simply having a bad day, which would have minimal impact on the existing knowledge about Pikachu. Thus, the decision to update or sustain the existing knowledge shapes how we process new social information.

Critically, the current study also examined the level of reactivation at the time of new encoding and assessed how it might be linked to the subsequent memory for the new, contrasting information. Our findings highlight that the specific role of reactivation in modulating updating depends on the strength of the existing memories. By measuring retrieval precision of an episodic detail from the existing memories during new learning, we inferred the level of reactivation using a behavioral metric. We show that reactivation during new learning only reflects memory strength when the incoming information conforms with expectations. However, when new information violates existing memories, weakly encoded memories are reactivated to a level equivalent to strong memories. Moreover, in both Experiments 1 and 2, we found that reactivation of the weak memories was significantly higher when a character’s emotion changed versus stayed the same (although this effect did not hold in the combined sample across all experiments), which may suggest that reactivation of existing memories can be triggered by the need to update prior knowledge⁸⁴. In contrast to the weakly encoded memories, the strong memories show a lower tendency to incorporate new information, and for them, remaining in a state with less plasticity (e.g., not reactivated) may help avoid unwanted changes. This may explain why the reactivation levels of the thrice studied Phase 1 events were not greater than, but equivalent to, those of the once studied events. Taken together, our results extend beyond prior work showing that memory reactivation can trigger plasticity and enable memory updating^4,5,23,25,26 and point to an intriguing possibility that memory reactivation and updating are not automatic but, instead, can be modulated by the need to update existing memories.

We also reveal that on a trial-by-trial basis, reactivation of strong memories precisely modulates their malleability upon encountering prediction errors, such that the likelihood of remembering the new information conflicting with a strong prior memory can be predicted by the level of reactivation for the given memory. This result is broadly consistent with prior work suggesting that reactivation during new learning predicts the integration of related memories^28,45,85,86. It is worth noting, however, in our study, reactivation only tracks the updating of individual strong memories, but not of the individual weak memories, and thus the generalizability of the relationship between memory reactivation and memory updating should be interpreted with caution.

Notably, very few studies have used behavioral metrics to index the levels of memory reactivation (but see⁸⁷ for a review on examining reactivation with eye movements). One innovation of the current study is probing the precision of the retrieved episodic detail as a way to assess reactivation of existing memories. This approach provides a continuous measure of a memory behavior supported by reactivation, introducing a potential tool for investigating reactivation processes in future behavioral studies. However, we acknowledge that the extent to which our measure reflects neural reactivation remains to be determined—a question that requires further evaluation using a paradigm combining behavioral and neuroimaging techniques. Future work seeking to adopt similar behavioral metrics to continuously track reactivation should also take task difficulty and potential data exclusions into consideration, to ensure that the assessed behavior sufficiently captures the underlying memory processes. It is also worth noting that, in our study, the tested shape information was an element peripheral to the memory violation experience. Therefore, it poses an open question of whether the reactivation of the shape information was correlated with the reactivation of other components from the same event, for instance and of particular interest, the emotional information. While existing work has provided evidence for memory dependency between different elements encoded within the same events^88,89, future investigation is needed to formally quantify the contingencies between their reactivation in the context of memory violation.

Notably, some of our results are also broadly consistent with the memory interference framework, such that the impaired memory of the new, conflicting emotional information in the prior strong as compared to prior weak condition may be due to proactive memory interference^{14,15,17,19,20}. Specifically, it would suggest that the corresponding Phase 1 and Phase 2 picnic events were represented as antagonistic memories, and that recalling the initial picnic events inhibits the retrieval of the new picnic events. Importantly, theoretical perspectives have suggested that this bias during retrieval can be modulated by the relative strength of the competing memories, such that there should be more interference from strong than weak memories^16–18. While this framework aligns with our result showing worse recall of the conflicting emotional information encoded subsequent to strong versus weak initial learning, it does not explain the null difference in the recognition memory for the specific Phase 2 GIFs between prior strong and prior weak conditions. In our study, the encoding strength of prior memories specifically influenced the retrieval of the emotional information, rather than interfering memory for all aspects of the related new events. This raises an open question of whether retrieval of the element that drives expectation violation in a new event – such as the change in emotional information in the current study – is particularly sensitive to the differences in existing memory strength. It is important to note, however, that the non-significant difference in GIF recognition accuracy may be partially attributed to reduced statistical power, as this test was only conducted in two of three experiments. Therefore, we acknowledge that based on the current study alone, it is hard to fully rule out memory interference playing a significant role in our findings. Existing perspectives have proposed potential factors that may bias conflicting events towards being stored as an integrated memory, as posited by the aforementioned updating framework, versus competing memories, as suggested by the interference framework; these factors include the degree of overlap across events and reactivation strength at the time of new learning^1,3,90. In the current study, participants learned largely overlapping picnic events that shared characters and food items, yet differed in emotional information. However, it remains hard to determine whether such degree of overlap favors memory integration versus competition. For similar reasons, although we also measured the level of reactivation during new encoding, our data are insufficient to fully dissociate these accounts. To further arbitrate between these frameworks, future studies may consider adopting neuroimaging approaches to characterize the memory representational structure of related events that contain conflicting information.

Additionally, although the current study primarily focuses on memories of different strengths when they are violated, our data also demonstrate memory benefits when incoming information conforms versus violates expectations. Participants showed significantly better performance recalling both Phase 1 and Phase 2 emotional information when the studied characters showed consistent versus opposite emotional valence towards the same food items, suggesting proactive and retroactive memory facilitation when the new information aligns with existing knowledge. This result is in line with prior findings that predictable information and congruency between experiences enhance memory^50,91–98. When considered together with findings in the emotion “change” condition, our results show that prior memory strength modulates new learning in opposite directions when the new information is congruent versus incongruent with existing knowledge (see Figure S2 left). Encoding strength of the existing memories positively predicts memories for the new experiences in the former case (when congruent), but negatively predicts them in the latter case (when incongruent). Previous framework has suggested a U-shaped function between prediction error and encoding, such that both consistency (i.e., null or low prediction errors) and surprise (i.e., high prediction errors) boosts learning^98–100; cf.,¹⁰¹. Here, our results add nuances to this framework by showing that expectation confirmation and violation differentially interact with existing memory strength to influence new learning.

While offering insights into the mechanisms underlying memory changes when violated, the findings reported in the current study are based on data from a large sample synthesized across three experiments. When examined within individual experiments with more typical sample sizes, however, some effects were only marginal or reached statistical significance in some of the experiments (see Supplemental Data for results from each individual experiment), which is a limitation of the current study. While we have carefully examined whether and how experiment as a factor might interact with the effects of our primary manipulations using a model comparison approach, we did not have a priori hypotheses regarding any experiment-specific effect and thus refrain from overinterpreting these interactions. Although speculative, we suspect that the divergence in some results between experiments may be attributed to small effect sizes, motivational differences between online participants, stage of COVID-19 quarantine or other pandemic-related factors. It would be valuable for future work to explore the boundary conditions allowing for the replication of our findings.

In conclusion, we show that in a social learning context, strongly encoded associative memories are more resistant to incorporating new information than weakly encoded memories. The likelihood for a specific strong memory to incorporate new information, however, is tracked by the degree of its reactivation at the time when a memory violation occurs. While this relationship is not present for specific weak memories, the overall level of reactivation for these weakly encoded memories is equivalent to that that of strong memories during counterfactual new learning. Collectively, our study highlights the adaptive feature of social memory updating by showing that when and how we remember conflicting information are contingent on the strength of prior related memories.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

W.Y. and L.D. designed research; W.Y. performed research and analyzed data; L.D. oversaw research and data analysis; and W.Y. and L.D. wrote the paper.

Funding Statement

This work was supported National Institutes of Health (NIH) Grant R01-MH0746492.

Data availability

We share de-identified data and analysis code on the Open Science Framework at https://osf.io/wy2zu.

Declarations

Competing interests

The authors declare no competing interests.