During rest and sleep, sequential neural activation patterns corresponding to awake experience re-emerge, and this replay has been shown to benefit subsequent behavior and memory. Whereas some studies show that replay directly recapitulates recent experience, others demonstrate that replay systematically deviates from the temporal structure, the statistics, and even the content of recent experience. Given these disparate characteristics, what is the nature and purpose of replay? Here, we offer a theoretical framework in which replay reflects simple context-guided processes that facilitate memory. We suggest that during awake learning, the brain associates experiences with the contexts in which they are encoded, at encoding rates that vary according to the salience of each experience. During quiescence, replay emerges as the result of a cascade of autonomous bidirectional interactions between contexts and their associated experiences, which in turn facilitates memory consolidation. A computational model instantiating this proposal explains numerous replay phenomena, including findings that existing models fail to account for and observations that have been predominantly construed through the lens of reinforcement learning. Our theory provides a unified, mechanistic framework of how the brain initially encodes and subsequently replays experiences in the service of memory consolidation.

Long-term learning and memory in the primate brain rely on a series of hierarchically organized subsystems extending from early sensory neocortical areas to the hippocampus. The components differ in their representational attributes and plasticity, with evidence for sparser, more decorrelated activity and faster plasticity in regions higher up in the hierarchy. How and why did the brain arrive at this organization? We explore the principles that allow such an organization to emerge by simulating a hierarchy of learning subsystems in artificial neural networks (ANNs) using a meta-learning approach. As ANNs optimized weights for a series of tasks, they concurrently meta-learned layer-wise plasticity and sparsity parameters. This approach enhanced the computational efficiency of ANNs, promoting hidden activation sparsity while benefitting task performance. Meta-learning also gave rise to a brain-like hierarchical organization, with higher layers displaying faster plasticity and a sparser, more pattern-separated neural code than lower layers. Early layers peaked early in their plasticity and stabilized, whereas higher layers continued to develop and maintained elevated plasticity over time, mirroring empirical developmental trajectories. Moreover, when trained on dual tasks imposing competing demands for item discrimination and categorization, ANNs with parallel pathways developed distinct representational and plasticity profiles, convergent with the distinct properties observed empirically across intra-hippocampal pathways. These results suggest that the macroscale organization and development of heterogeneous learning subsystems in the brain may emerge in part from optimizing biological variables that govern plasticity and sparsity. ### Competing Interest Statement The authors have declared no competing interest.

Memories for negative life experiences play a key role in our mental health and are processed by the sleeping brain. Here, we attempted to bias the sleeping brain towards processing a more adaptive version of a negative autobiographical memory during a nap using targeted memory reactivation (TMR). Forty-seven participants (23.6±5.5 years old; 68% female) provided details about one negative and one neutral life experience and provided self-report ratings on arousal to each memory (0=very calm; 10=very aroused) and distress to the dominant emotion associated with their negative memory (0=no distress; 10=extreme distress). They completed an imagery rescripting intervention for their negative memory during which they heard a repetitive sound cue and a similar procedure for their neutral memory during which they heard a different cue. All participants were then allowed a 2-hour PSG-monitored nap opportunity in the lab and were randomized to hear either the sound cue associated with their negative (n=22) or neutral (n=18) memory (i.e., to have one memory “targeted” for reactivation); participants who got less than 15 minutes of non-REM sleep were reclassified into a “wake” group (n=7). After the nap and one week later, participants repeated arousal and distress ratings. Arousal to both negative and neutral memories decreased after the interventions and nap period (β=-1.64, p< 0.0001) and remained lower one week later (β=-1.51, p< 0.0001), but this did not depend on whether the memory was cued via TMR. Similarly, emotional distress decreased after imagery rescripting (β=-2.98, p< 0.0001) and decreased further after the nap period (β=-1.00, p=0.04), but this reduction did not depend on TMR cueing. Participants in the wake group appeared to show a different pattern in which distress to dominant emotions increased after the nap period, but this difference was not significant in the current sample. TMR did not enhance the arousal- and distress-reducing effects of imagery rescripting, but a nap did appear to further reduce emotional distress. Further research on the benefits of sleep and TMR on processing negative autobiographical memories is warranted.

Concepts contain rich structures that support flexible semantic cognition. These structures can be characterized by patterns of feature covariation: certain clusters of features tend to occur in the same items (e.g., feathers, wings, can fly). Existing computational models demonstrate how this kind of structure can be leveraged to slowly learn the distinctions between categories, on developmental timescales. It is not clear whether and how we leverage feature structure to quickly learn a novel category. We thus investigated how the internal structure of a new category is extracted from experience and what kinds of representations guide this learning. We predicted that humans can leverage feature clusters within an individual category to benefit learning and that this relies on the rapid formation of distributed representations. Novel categories were designed with patterns of feature associations determined by carefully constructed graph structures (Modular, Random, and Lattice). In Experiment 1, a feature inference task using verbal stimuli revealed that Modular categories—containing clusters of reliably covarying features—were more easily learned than non-Modular categories. Experiment 2 replicated this effect using visual categories. In Experiment 3, a temporal statistical learning paradigm revealed that this Modular benefit persisted even when category structure was incidental to the task. We found that a neural network model employing distributed representations was able to account for the effects, whereas prototype and exemplar models could not. The findings constrain theories of category learning and of structure learning more broadly, suggesting that humans quickly form distributed representations that reflect coherent feature structure.

Our environment contains temporal information unfolding simultaneously at multiple timescales. How do we learn and represent these dynamic and overlapping information streams? We investigated these processes in a statistical learning paradigm with simultaneous short and long timescale contingencies. Human participants (n = 96) played a game where they learned to quickly click on a target image when it appeared in one of nine locations, in eight different contexts. Across contexts, we manipulated the order of target locations: at a short timescale, the order of pairs of sequential locations in which the target appeared; at a longer timescale, the set of locations that appeared in the first versus the second half of the game. Participants periodically predicted the upcoming target location, and later performed similarity judgments comparing the games based on their order properties. Participants showed context-dependent sensitivity to order information at both short and long timescales, with evidence of stronger learning for short timescales. We modeled the learning paradigm using a gated recurrent network trained to make immediate predictions, which demonstrated multilevel learning timecourses and patterns of sensitivity to the similarity structure of the games that mirrored human participants. The model grouped games with matching rule structure and dissociated games based on low-level order information more so than high-level order information. The work shows how humans and models can rapidly and concurrently acquire order information at different timescales.

Our representations of the world need to be stable enough to support general knowledge but flexible enough to incorporate new information as our environment changes. How does the human brain manage this stability-plasticity trade-off? We analyzed a large dataset in which participants viewed objects embedded in thousands of natural scenes across many fMRI sessions. Semantic item representations were located by jointly leveraging a voxelwise encoding model to find reliable item representations and a word-embedding model to evaluate semantic content. Within the medial temporal lobe, semantic item representations in hippocampal subfield CA1, parahippocampal cortex, and perirhinal cortex gradually drifted across a period of multiple months. Whole-brain analyses revealed a gradient of plasticity in the temporal lobe, with drift more evident in anterior than posterior areas. On short timescales, rapid plasticity was observed only in parahippocampal cortex, such that item co-occurrence statistics warped item representations within a single session. Together, the results suggest that the brain solves the stability-plasticity trade-off through a gradient of plasticity across semantic regions.

Memory reactivation during sleep is thought to facilitate memory consolidation. Most sleep reactivation research has examined how reactivation of specific facts, objects, and associations benefits their overall retention. However, our memories are not unitary, and not all features of a memory persist in tandem over time. Instead, our memories are transformed, with some features strengthened and others weakened. Does sleep reactivation drive memory transformation? We leveraged the Targeted Memory Reactivation technique in an object category learning paradigm to examine this question. Participants (20 female, 14 male) learned three categories of novel objects, where each object had unique, distinguishing features as well as features shared with other members of its category. We used a real-time EEG protocol to cue the reactivation of these objects during sleep at moments optimized to generate reactivation events. We found that reactivation improved memory for distinguishing features while worsening memory for shared features, suggesting a differentiation process. The results indicate that sleep reactivation does not act holistically on object memories, instead supporting a transformation process where some features are enhanced over others.

A theory and neurocomputational model are presented that explain grid cell responses as the byproduct of equally dissimilar hippocampal memories. On this account, place and grid cells are not best understood as providing a navigational system. Instead, place cells represent memories that are conjunctions of both spatial and non-spatial attributes, and grid cells primarily represent the non-spatial attributes (e.g., odors, surface texture, etc.) found throughout the two-dimensional recording enclosure. Place cells support memories of the locations where non-spatial attributes can be found (e.g., positions with a particular odor), which are arranged in a hexagonal lattice owing to memory encoding and consolidation processes (pattern separation) as applied to situations in which the non-spatial attributes are found at all locations of a two-dimensional surface. Grid cells exhibit their spatial firing pattern owing to feedback from hippocampal place cells (i.e., a hexagonal pattern of remembered locations for the non-spatial attribute represented by a grid cell). The model explains: 1) grid fields that appear to be centered outside the box; 2) the toroidal nature of grid field representations; 3) grid field alignment with the enclosure borders; 4) modules in which grid cells have the same orientation and spacing but different phases; 5) head direction conjunctive grid cells that become simple head direction cells in the absence of hippocampal feedback; 6) the instant existence of grid fields in a novel environment; 7) the slower learning of place cells; 8) the manner in which head direction sensitivity of place cells changes near borders and in narrow passages; 9) the kinds of changes that underlie remapping of place cells; and 10) grid-like responses for two-dimensional coordinate systems other than navigation.

Some information links our experiences together while other information sets them apart. This poses a challenge for our memory systems, as learning shared features benefits from integration across instances to capture similarities, whereas learning unique features benefits from separation to avoid interference. We leveraged a color memory distortion paradigm to evaluate how we approach this representational tension when rapidly learning a structured novel domain. In two experiments, we trained participants over the course of half an hour on the shared and unique features of categories of novel objects, where each feature had a color drawn from a 2D continuous color space. There were no differences in how accurately participants remembered the color of shared and unique features overall, but when inaccurate, participants misremembered the color of shared (relative to unique) features as more similar to the category’s average color, suggesting that these features become more integrated. We show the same rapid representational warping in a neural network model trained on the same categories, providing an account of how a simple learning process can generate these biases. Together, this work shows how memories for different features, even those within the same object, become rapidly and differentially warped as a function of their roles in a category.

Research on the role of the hippocampus in memory acquisition has generally focused on active learning. But to understand memory, it is at least as important to understand processes that happen offline, during both wake and sleep. In a study of patients with amnesia, we previously demonstrated that although a functional hippocampus is not necessary for the acquisition of procedural motor memory during training session, it is required for its offline consolidation during sleep. Here, we investigated whether an intact hippocampus is also required for the offline consolidation of procedural motor memory while awake. Patients with amnesia due to hippocampal damage (n = 4, all male) and demographically matched controls (n = 10, 8 males) trained on the finger tapping motor sequence task. Learning was measured as gains in typing speed and was divided into online (during task execution) and offline (during interleaved 30 s breaks) components. Amnesic patients and controls showed comparable total learning, but differed in the pattern of performance improvement. Unlike younger adults, who gain speed across breaks, both groups gained speed only while typing. Only controls retained these gains over the breaks; amnesic patients slowed down and compensated for these losses during subsequent typing. In summary, unlike their peers, whose motor performance remained stable across brief breaks in typing, amnesic patients showed evidence of impaired access to motor procedural memory. We conclude that in addition to being necessary for the offline consolidation of motor memories during sleep, the hippocampus maintains access to motor memory across brief offline periods during wake.

A core feature of episodic memory is representational drift, the gradual change in aggregate oscillatory features that supports temporal association of memory items. However, models of drift overlook the role of episodic boundaries, which indicate a shift from prior to current context states. Our study focuses on the impact of task boundaries on representational drift in the parietal and temporal lobes in 99 subjects during a free recall task. Using intracranial EEG recordings, we show boundary representations reset gamma band drift in the medial parietal lobe, selectively enhancing the recall of early list (primacy) items. Conversely, the lateral temporal cortex shows increased drift for recalled items but lacked sensitivity to task boundaries. Our results suggest regional sensitivity to varied contextual features: the lateral temporal cortex uses drift to differentiate items, while the medial parietal lobe uses drift-resets to associate items with the current context. We propose drift represents relational information tailored to a region’s sensitivity to unique contextual elements. Our findings offer a mechanism to integrate models of temporal association by drift with event segmentation by episodic boundaries.

Sleep plays an important role in memory processing and is disrupted in individuals with post-traumatic stress disorder (PTSD). A growing body of research has experimentally investigated how sleep - or lack thereof - in the early aftermath of a traumatic experience contributes to intrusive memory formation. The aim of this meta-analytic review was to examine the effects of various experimental sleep manipulations (e.g., sleep deprivation, daytime naps) on intrusive memories following exposure to an experimentally induced analogue traumatic event. Eight eligible studies were systematically identified through PsycInfo and PubMed and provided sufficient data to contribute to a meta-analysis of the effects of sleep versus wakefulness on intrusive memory frequency. Sleep was found to reduce intrusive memory frequency when compared to wakefulness at a small but significant effect size (Hedge's g = 0.29). There was no evidence of publication bias and heterogeneity of effect sizes across studies was moderate. Results suggest that sleep plays a protective role in the aftermath of exposure to a traumatic event with implications for early post-trauma intervention efforts.

We are unresponsive during slow-wave sleep but continue monitoring external events for survival. Our brain wakens us when danger is imminent. If events are non-threatening, our brain might store them for later consideration to improve decision-making. To test this hypothesis, we examined whether novel vocabulary consisting of simultaneously played pseudowords and translation words are encoded/stored during sleep, and which neural-electrical events facilitate encoding/storage. An algorithm for brain-state dependent stimulation selectively targeted word pairs to slow-wave peaks or troughs. Retrieval tests were given 12 and 36 hours later. These tests required decisions regarding the semantic category of previously sleep-played pseudowords. The sleep-played vocabulary influenced awake decision-making 36 hours later, if targeted to troughs. The words' linguistic processing raised neural complexity. The words' semantic-associative encoding was supported by increased theta power during the ensuing peak. Fast-spindle power ramped up during a second peak likely aiding consolidation. Hence, new vocabulary played during slow-wave sleep was stored and influenced decision-making days later.

Full text Figures and data Side by side Abstract eLife assessment Introduction Results Discussion Materials and methods Data availability References Peer review Author response Article and author information Abstract Maintaining an accurate model of the world relies on our ability to update memory representations in light of new information. Previous research on the integration of new information into memory mainly focused on the hippocampus. Here, we hypothesized that the angular gyrus, known to be involved in episodic memory and imagination, plays a pivotal role in the insight-driven reconfiguration of memory representations. To test this hypothesis, participants received continuous theta burst stimulation (cTBS) over the left angular gyrus or sham stimulation before gaining insight into the relationship between previously separate life-like animated events in a narrative-insight task. During this task, participants also underwent EEG recording and their memory for linked and non-linked events was assessed shortly thereafter. Our results show that cTBS to the angular gyrus decreased memory for the linking events and reduced the memory advantage for linked relative to non-linked events. At the neural level, cTBS targeting the angular gyrus reduced centro-temporal coupling with frontal regions and abolished insight-induced neural representational changes for events linked via imagination, indicating impaired memory reconfiguration. Further, the cTBS group showed representational changes for non-linked events that resembled the patterns observed in the sham group for the linked events, suggesting failed pruning of the narrative in memory. Together, our findings demonstrate a causal role of the left angular gyrus in insight-related memory reconfigurations. eLife assessment This important paper provides solid evidence that the angular gyrus plays a role in insight-based memory updating. The study is well conducted, timely, and presents clear-cut behavioral results. While the study provides robust evidence that transcranial magnetic stimulation to the angular gyrus impacts memory, evidence for the strong claim of a causal contribution of the angular gyrus in particular – apart from other connected regions, including the hippocampus – is not conclusive. https://doi.org/10.7554/eLife.91033.3.sa0 About eLife assessments Introduction The capacity to flexibly update our memories in light of new information is fundamental to maintaining an accurate model of the world around us. This flexibility requires adaptable memory networks that can be reconfigured upon acquiring new insights. Previous research provided direct evidence for insight-induced reconfigurations of memory representations and showed that insight into the connection of initially separate events propels the integration of these events into coherent episodes (Collin et al., 2015; Milivojevic et al., 2015). Such mnemonic integration allows novel inferences (Spalding et al., 2018; Zeithamova et al., 2012) that aid efficient navigation (Coutanche et al., 2013; Fernandez et al., 2023; He et al., 2022) and decision-making (Boorman et al., 2021; Kumaran et al., 2009; Shohamy and Daw, 2015). Importantly, in everyday life, the inference about the relationship between seemingly unrelated events is often not inferred via direct observation but through imagination. For instance, when reading a book, we gain insight into the plot and possible twists through our imagination, which then prompts us to update our memory representations. Even when new insights are derived from direct observation, the integration process requires imaginative capacities to bind the previously separate memories into a coherent narrative. At the neural level, the hippocampus has been shown to play a pivotal role in (imagination-based) mnemonic integration (Cohn-Sheehy et al., 2021a; Collin et al., 2015; Griffiths and Fuentemilla, 2020; Grob et al., 2023a; Milivojevic et al., 2015). However, while the hippocampus appears to be crucial for mnemonic integration, it does not act in isolation but operates in collaboration with cortical areas to accomplish this complex process (Backus et al., 2016; Milivojevic et al., 2015; Pehrs et al., 2018; Schlichting and Preston, 2015; Spalding et al., 2018). Yet, our understanding of the specific areas implicated in the insight-driven reconfiguration of memory representations, beyond the hippocampus, remains limited. Moreover, existing data on the neural underpinnings of mnemonic integration are mainly correlational in nature and which areas are causally involved in the integration of initially unrelated memories into cohesive representations is completely unknown. One promising candidate that may contribute to insight-driven memory reconfiguration is the angular gyrus. The angular gyrus has extensive structural and functional connections to many other brain regions (Petit et al., 2023), including the hippocampus (Coughlan et al., 2023; Uddin et al., 2010). Accordingly, previous studies have shown that stimulation of the angular gyrus resulted in altered hippocampal activity (Thakral et al., 2020; Wang et al., 2014). Furthermore, the angular gyrus has been implicated in a myriad of cognitive functions, including mental arithmetic, visuospatial processing, inhibitory control, and theory-of-mind (Cattaneo et al., 2009; Grabner et al., 2009; Lewis et al., 2019; Schurz et al., 2014). Moreover, there is accumulating evidence pointing to a key role of the angular gyrus in long-term memory (Bellana et al., 2017; Bonnici et al., 2018; Kwon et al., 2022; Wang et al., 2014) and imagination (Ramanan et al., 2018; Thakral et al., 2017; Thakral et al., 2020). How these putative functions of the angular gyrus relate to one another, however, remained unclear. We reasoned that these functions might be directly linked, enabling the angular gyrus to drive the integration of (imagination-related) insights into long-term memory. In line with this idea, recent theories propose that the angular gyrus acts as dynamic buffer for spatiotemporal representations (Humphreys et al., 2021), which may allow the angular gyrus to transiently maintain the initially separate events and to integrate these into cohesive narratives. This buffering function of the angular gyrus may be particularly relevant for imagination-based linking. Thus, we hypothesized that the angular gyrus plays a crucial role in integrating imagination-related insights into long-term memory and hence in the dynamic reconfiguration of memory representations in light of new information. To test this hypothesis and determine the causal role of this area in insight-related memory reconfigurations, we conducted a preregistered study combining a life-like video-based narrative-insight task (Milivojevic et al., 2015; Figure 1), probing insight-related reconfigurations of memory, with representational similarity analysis of EEG data and (double-blind) ‘neuro-navigated’ TMS to an area of the left angular gyrus that was implicated in imaginative processing before (Thakral et al., 2017). Considering this involvement of the angular gyrus in imaginative processes, we expected that the effect of cTBS on the change in representational similarity from pre- to post-insight will differ based on the mode of insight – whether this insight was gained via imagination or observation. Specifically, we expected a more pronounced impairment in the neural reconfigurations when insight is gained via imagination, as this function may depend more on angular gyrus recruitment than insight gained via observation. Additionally, we expected cTBS to the left angular gyrus to reduce the increase in neural similarity for linked events and increase of neural dissimilarity for non-linked events. We further predicted that cTBS to the left angular gyrus would specifically reduce the impact of (imagination-based) insight into the link of initially unrelated events on memory performance during free recall, given the higher variability of free recall compared to other memory measures with lower search demands. Considering the high connectivity profile of the angular gyrus within the brain (Seghier, 2013), we conducted an EEG connectivity analysis building upon findings from the RSA analyses concerning alterations in neural reconfigurations. To establish a link between neural and behavioral findings, we chose a correlational approach to relate observations from these two domains. We intentionally adopted a mixed design, combining both between-subjects and within-subject methodologies. The between-subjects approach was chosen to minimize the risk of carry-over effects and sequence biases. Simultaneously, we capitalized on the advantages of a within-subject design by altering the pre- to post-insight comparison and the mode of insight (imagination vs. observation) within each participant. To control for any group differences beyond the TMS manipulation, we gathered various control variables through questionnaires, including trait- and state-anxiety, depressive symptoms, chronic stress levels, personality dimensions, and imaginative capacities. Figure 1 Download asset Open asset Modified narrative-insight task and procedure. During the pre-phase, participants viewed video events (A, B, and X) from 10 different storylines. Each event was preceded by a title (1 s) and repeated 18 times. The inter-stimulus interval (ISI) was ~1000ms. The subsequent insight-phase consisted of two parts. In one part, participants gained insight through a written imagination instruction (I) interspersed with a control instruction (CI). In the other part, they gained insight through a linking video (L) interspersed with a control video (C). The order of gaining insight through imagination or video observation was counterbalanced across participants. Before each insight part, participants received, depending on the experimental group, either a sham or cTBS stimulation over the left angular gyrus (MNI: −48,–67, 30). After the insight-phase, participants had a 30-min break and then completed a free recall for a maximum of 20 min in a different room. In the post-phase, all video events were presented in the same manner as the pre-phase. Results cTBS to the angular gyrus reduces insight-related memory boost The angular gyrus has been implicated in a myriad of tasks and functions, including long-term memory (Bonnici et al., 2018; Kwon et al., 2022; Wang et al., 2014) and imagination (Ramanan et al., 2018; Thakral et al., 2017). Here, we hypothesized that these functions of the angular gyrus are directly linked to one another. Specifically, we postulated that the angular gyrus plays a crucial role in the integration of imagination-related insights into long-term memory representations and that it thus represents a key player in the dynamic reconfiguration of memory in light of new information. To test this hypothesis and the causal role of the angular gyrus in insight-related memory reconfigurations, we combined the life-like video-based narrative-insight task with representational similarity analysis of EEG data and (double-blind) neuro-navigated TMS over the left angular gyrus in a comprehensive investigation within a single day. During the narrative-insight task, participants first saw three video events (A, B, and X; pre-phase), which were then either linked into a narrative (A and B) or not (A and X) in a subsequent insight-phase. Critically, before the insight-phase, we applied either sham stimulation (31 participants, 15 females) or continuous theta burst stimulation (cTBS; 34 participants, 16 females) to the left angular gyrus. Notably, the groups did not differ on levels of subjective chronic stress (TICS), state and trait anxiety (STAI-S, STAI-T), depressive mood (BDI), imaginative capacities (FFIS), personality dimensions (BFI), age, and motor thresholds (for descriptive statistics see Table 1; all p>0.056). Table 1 Control variables. ShamcTBSMeasureMSDMSDpuncorrFFIS-C24.064.2423.594.640.066FFIS-D16.845.4217.865.220.446FFIS-E13.847.1615.007.100.514FFIS-F27.878.9628.219.630.885STAI-T34.139.2738.6211.150.082STAI-S35.357.6039.6510.080.056TICS11.978.5413.419.950.531BDI6.846.887.657.570.654BFI-2 E43.105.6640.598.540.165BFI-2 N27.136.9430.1210.090.166BFI-2 O47.166.9946.686.890.779BFI-2 C40.428.0640.716.780.878BFI-2 A48.945.0646.745.760.106Age25.454.6223.623.820.088MT53.0314.5954.8212.930.608 Note. The questionnaires FFIS with its dimensions: FFIS-C (complexity of imagination), FFIS-D (directedness of imagination), FFIS-E (emotional valence of imagination), FFIS-F (frequency of imagination); STAI-T and STAI-S; TICS; BDI; BFI-2 with its dimensions: BFI-2 E (extraversion), BFI-2 N (neuroticism), BFI-2 O (openness to experience), BFI-2 C (conscientiousness), BFI-2 A (agreeableness) were completed during the 30-min break after the insight-phase. Age in years. Motor thresholds (MT) in percent of maximum stimulator capacity. No significant group differences were observed on any of these measures. p Values are displayed uncorrected for multiple comparisons. Data represents means (+/-SD). Following the insight-phase and a 30-min break to mitigate potential TMS aftereffects (Huang et al., 2005; Jannati et al., 2023), participants completed a free recall task, which provided a measure of insight-related changes in subsequent memory. Thereafter, participants saw the same video events (A, B, and X) again in a post-phase. EEG was measured during all stages of the narrative-insight task. Contrasting neural representation patterns from the pre- and post-phases allowed us to assess insight-related memory reconfiguration and its modulation by cTBS to the angular gyrus. Due to its specific relevance in imaginative processes (Ramanan et al., 2018; Thakral et al., 2017; Thakral et al., 2020), we expected that the angular gyrus would be particularly relevant if insight relies strongly on imagination. Therefore, participants gained insight into half of the stories by imagining the link themselves, while they observed the link as a video in the other half of the stories. Participants’ ratings showed that they adhered well to these instructions during the linking phase. When linking events via imagination, they reported imagining the linking events very well (M = 3.38, SD = 0.47) and their imagination as depictive (M = 3.35, SD = 0.46). When linking via observation, they reported a high level of understanding of the linking events (M = 3.37, SD = 0.51) and found the linking events meaningful (M = 3.35, SD = 0.52) on a 1–4 Likert scale. Furthermore, participants demonstrated a high level of attention throughout the narrative-insight task, responding to target stimuli with near-ceiling performance (M = 99.25%; SD = 1.40 %) without any group differences (t(63.00) = 0.42, p = 0.675, d = –0.10). Importantly, participants were unaware of the allocation to the cTBS or sham condition, as indicated by the treatment guess at the end of the experiment (Fisher’s exact test; p = 0.597). Furthermore, TMS stimulation did not affect participants’ subjective mood, wakefulness or arousal (mood: group ×time: F(1, 63) = 0.76, p = 0.386, ηGG0.00; wakefulness: group ×time: F(1, 63) = 0.01, p = 0.921, ηGG0.00; arousal: group ×time: F(1, 63) = 0.01, p = 0.921, ηGG0.00). As expected, all participants gained insight into which events were linked in the narrative-insight task, as they rated the belongingness of linked events higher than non-linked events from pre- to post-insight, as indicated by a linear mixed model (LMM: time ×link: β = 2.49, 95% CI [2.15, 2.83], t(418.44) = 14.01, p<0.001; Figure 2—figure supplement 1). Post-hoc tests showed increasing belongingness ratings for linked events and decreasing belongingness ratings for non-linked events from pre- to post-insight (LMM: link: β = 1.49, 95% CI [1.33, 1.64], t(418) = 24.49, p<0.001; non-link: β = –.91, 95% CI [–1.07, –.76], t(418.00) = –15.03, p<0.001). This insight was further reflected in the multi-arrangements task (MAT), in which participants were instructed to arrange representative images (A, B, and X) from each story based on their relatedness. In this task, all participants arranged linked events closer together than non-linked events (MAT; LMM: link: β = –1.33, 95% CI [-1.59,–1.07], t(177.00) = –9.81, p<0.001; Figure 2—figure supplement 2). The strong insight gained by all participants was further reflected in their near-ceiling performance in the forced-choice recognition task, in which participants were instructed to identify the event (B or X) that was linked with A. Participants accurately indicated whether B or X was linked to A (sham: M = 94.65%, SD = 9.00%; cTBS: M = 97.34%, SD = 6.10%; Figure 2—figure supplement 3). Importantly, there were no group differences in any of these measures (LMM: narrative-insight task: group × time × link: β = –0.02, 95% CI [–0.49, 0.45], t(418.27) = –0.09, p = 0.929; LMM: multi-arrangements task: group × link: β = 0.11, 95% CI [–0.26, 0.48], t(177.00) = 0.58, p = 0.561; LMM: Forced-choice recognition: group: β = 0.23, 95% CI [–0.26, 0.72], t(113.37) = 0.90, p = 0.368), indicating that all participants successfully gained insight into which events were linked and that the (left) angular gyrus did not play a critical role in the process of gaining insight itself. To investigate the causal role of the left angular gyrus in insight-related episodic memory integration, the key question of this study, we first analyzed the detailedness of participants’ memory for both linked and non-linked events during free recall. Across groups, linked events were generally recalled in more detail than non-linked events (LMM: link: β = 1.20, 95% CI [0.86, 1.54], t(406.00) = 6.75, p<0.001), suggesting a memory boost for integrated narratives. Most interestingly, cTBS to the left angular gyrus through cTBS reduced this insight-related memory boost for linked events significantly (LMM: group × link: β = –0.54, 95% CI [-1.02,–0.06], t(406.00) = –2.17, p = 0.030; Figure 2A). Pairwise comparisons revealed a significantly lower number of recalled details for linked events in the cTBS compared to the sham group, while there was no significant difference for non-linked events (LMM: link: β = –0.40, 95% CI [-0.79,–0.02], t(85.50) = –2.79, p = 0.033; LMM: non-link: β = –0.09, 95% CI [–0.29, 0.42], t(406.00) = –0.62, p = 0.926). Additionally, we observed that all participants showed better memory for central compared to peripheral details of the plot when recalling linked events, which was not observed to the same extent for non-linked events (LMM: link × detail: β = 0.61, 95% CI [0.12, 1.09], t(406.00) = 2.43, p = 0.016). Figure 2 with 3 supplements see all Download asset Open asset Behavioral results. (A) Significantly reduced recall of details for linked events in the cTBS group compared to the sham group, with no significant difference for the non-linked events. (B) Significantly reduced recall of the linking events in the cTBS group compared to the sham group. (C) Schematic overview of electric field modeling: Simulation was performed for the angular gyrus coordinate (MNI: x = –48, y = –67, z = 30) using a Magstim 70 mm figure-of eight coil at 80% of individual motor thresholds, reflecting the applied setup. The resulting electric field was averaged within a 10 mm spherical ROI and centered on the target coordinate and extracted for subsequent analyses. Please note that in the study, the coil handle was oriented upwards; however, in this illustration, it has been intentionally depicted as pointing downwards for better visibility purposes. (D) Significantly reduced number of details recalled for linked events specifically in the high cTBS group (based on a median-split on simulated electric field strengths). (E) Significantly reduced recall of the linking events specifically in the high cTBS group (based on a median-split on simulated electric field strengths). Boxplots show the median for each group. Boxplot whiskers extend to the minimum or maximum value within 1.5 times the interquartile range. Points within the boxplot indicate individual data points per each group. Density plots indicate data distribution per group. The belongingness ratings for the linked and non-linked events are shown in Figure 2—figure supplement 1, the data of the multiple arrangements task in Figure 2—figure supplement 2, and the data of the forced-choice recognition test in Figure 2—figure supplement 3. Statistical differences stem from pairwise post-hoc tests of marginal means. *p<0.05, ***p<0.001. In a second step, we analyzed whether cTBS to the angular gyrus affected, in addition to memory detailedness for initially separate but now linked events, also the memory for the linking events themselves. Our results showed that cTBS (vs. sham) significantly reduced the frequency with which participants recalled the linking events (LMM: group: β = –0.66, 95% CI [-1.13,–0.18], t(98.13) = –2.71, p = 0.008; Figure 2B). Interestingly, this TMS effect appeared to be particularly pronounced when events were linked via imagination (cTBS vs. sham: t(61.46) = –2.53, p = 0.014, d = –0.63) and was less prominent when they were linked via direct observation (cTBS vs. sham: t(58.59) = - 1.63, p = 0.107, d = –0.40), although it is important to note that the interaction was not significant (LMM: group × mode: β = 0.30, 95% CI [–0.18, 0.77], t(62) = 1.23, p = 0.225). To assess the effect of cTBS stimulation on the angular gyrus (Pizem et al., 2022; Zhang et al., 2022), we performed electric field simulations at 80% of the individual motor threshold, averaging the estimated field strength within a 10 mm sphere centered around the angular gyrus coordinate (MNI: −48,–67, 30). In order to examine whether the behavioral effects were dependent on the simulated electric field strength (Figure 2C), we next included electric field strength (strong vs. weak via median split) and repeated the previous linear mixed model predicting the number of details for linked events including a group factor reflecting stimulation strength (sham, low, high). This model yielded a significant group × link interaction (LMM: β = –0.78, 95% CI [-1.35,–0.21,], t(399.00) = –2.63, p = 0.009; Figure 2D), suggesting a dependency of memory on stimulation strength. Pairwise comparisons for linked events confirmed that a stronger electric field induction in the angular gyrus significantly reduced the memory boost for linked events, while there was no such effect for weak cTBS stimulation (LMM: sham vs. low: β = 0.05, 95% CI [–0.45, 0.55], t(87.70) = 0.28, p = 1.000; sham vs. high: β = 0.74, 95% CI [0.25, 1.23], t(87.7) = 4.45, p<0.001, low vs. high: β = 0.69, 95% CI [0.13, 1.26], t(87.7) = 3.60, p = 0.007). We further included the electric field strength (strong vs. weak via median split) and repeated the previous linear mixed model predicting the naming of the linking events including the group factor stimulation strength (sham, low, high). This analysis yielded a significant effect of group (LMM: β = –0.92, 95% CI [-1.50,–0.35], t(97.79) = –3.11, p = 0.003; Figure 2E), suggesting that the memory for the linking events was dependent on the angular gyrus stimulation strength. Angular gyrus stimulation disrupts neural pattern reconfiguration following imagination-based insight Our behavioral data showed that cTBS to the angular gyrus reduced the insight-related memory boost. In a next step, we tested whether cTBS to the angular gyrus may also alter the insight-related reconfiguration of neural memory representations, taking the mode of insight (i.e. imagination vs. observation) into account. To this end, we leveraged representational similarity analysis (RSA) of EEG data and compared changes in multivariate oscillatory theta power patterns for linked and non-linked events from pre- to post-insight (Figure 3A). We focused exclusively on the theta band since theta has been shown to hold a key role in episodic memory integration (Backus et al., 2016; Nicolás et al., 2021). For this analysis, similarity maps (time × time) were computed by correlating story-specific theta frequency patterns within linked (A with B) and within non-linked (A with X) events in the pre- and post-phase, separately. Subsequently, we examined insight-induced effects on neural representations for linked (vs. non-linked) events by comparing the change from pre- to post-insight (post-pre) and the difference between imagination and observation (imagination - observation) between cTBS and sham groups using an independent cluster-based permutation t-test. Figure 3 with 2 supplements see all Download asset Open asset Representational pattern changes. (A), Conceptual overview of the representational similarity analysis (RSA) on theta oscillations. First, time-frequency data was computed, and the theta power values (4–7 Hz) were extracted. Using these feature vectors, Pearson’s correlations were computed to compare the power patterns across time points of events (here: event A and B). These correlations resulted in a time × time similarity map. (B), Significant cluster, denoted by white dotted line for illustrative purposes, for the change from post-pre and imagination-observation between the cTBS and sham groups using an independent sample cluster-based permutation t-test for linked events (A and B). In the middle panel, follow-up tests on stories linked via imagination revealed increased similarity for the sham group, while no significant effect was observed for the cTBS group. In the lower panel, follow-up tests on stories linked via observation showed decreased similarity for the sham group and increased similarity for the cTBS group. (C), Significant cluster, denoted by white dotted line for illustrative purposes, for the change from post-pre and imagination-observation between the cTBS and sham groups using an independent sample cluster-based permutation t-test for non-linked events (A and X). In the middle panel, follow-up tests on stories linked via imagination revealed increased similarity for the cTBS group, while no significant effect was observed for the sham group. In the lower panel, follow-up tests on stories linked via observation showed decreased similarity for the cTBS group and no significant effect for the sham cTBS group. Boxplots show the median similarity for each group at each time point. Boxplot whiskers extend to the minimum or maximum value within 1.5 times the interquartile range. Points within the boxplot indicate individual data points in each group. Density plots indicate data distribution per group and time. *p<0.05, **p<0.01, ***p<0.001. First, we included the within-subject factors time (pre vs. post), mode of insight (imagination vs. observation) and link (vs. non-link) by calculating the difference waves. Subsequently we conducted a cluster-based permutation test comparing the cTBS and the sham groups. This analysis yielded a four-way interaction within a negative cluster in a fronto-temporal region (electrode: FT7; p = 0.007, ci-range = 0.00, SD = 0.00). This result indicates that the impact of cTBS over the angular gyrus on the neural pattern reconfiguration following imagination- vs. observation-based insight may differ between linked and non-linked events. For linked events, this analysis yielded a negative cluster (p = 0.032, ci-range = 0.00, SD = 0.00) in the parieto-temporal region (electrodes: T7, Tp7, P7; Figure 3B; Figure 3—figure supplement 1). Follow-up tests on the extracted similarity cluster analyzed the representational pattern change and its modulation by TMS separately for the imagination and observation condition. For stories linked via imagination, we obtained an increase in representational similarity from pre- to post-insight in the sham group (t(30) = 3.48, p = 0.002, drepeated measures = 0.62), whereas there was no such increase and even a trend for a decrease in representational similarity for linked events from pre- to post-insight in the cTBS group (t(30) = –2.01, p = 0.053, drepeated measures = –0.36; group × time: F(1, 60) = 14.03, p<0.001, ηG = 0.09; Figure 3B middle panel). Interestingly, we observed that a lower change (post - pre) in representational similarity of events linked via imagination (vs. observation) was associated, across groups, with a reduced probability of recall of the linking events (r = 0.27, t(59) = 2.17, p = 0.034), suggesting a direct association between neural pattern reconfiguration and subsequent memory. Furthermore, to address a deviation from the normality assumption, the correlational analysis was repeated using the Spearman method, which indicated a stronger correlation (r(59) = 0.32, p = 0.012). For stories that were linked via observation, we observed a seemingly opposite pattern (group × time: F(1, 60) = 19.21, p<0.001, ηG = 0.12): decreased similarity in the sham group (t(30) = –3.94, p<0.001, drepeated measures = –0.62) but increased representational similarity in the cTBS group (t(30) = 2.30, p = 0.029, drepeated measures = 0.62; Figure 3B lower panel). However, these changes in representational similarity for the observation condition should be interpreted with caution, as these seemingly opposite changes appeared to be at least in part driven by group differences already in the pre-phase, before participants gained insight. Interestingly, we observed a different pattern of insight-related representational pattern changes for non-linked events. Similarly to linked events, we compared the change from pre- to post-insight and the difference between imagination and observation between cTBS and sham using an independent sample cluster-based permutation t-test. This analysis yielded a positive cluster (p = 0.035, ci-range = 0.

Declarative memory retrieval is thought to involve reinstatement of the neuronal activity patterns elicited and encoded during a prior learning episode. Recently, it has been suggested that two mechanisms operate during reinstatement, dependent on task demands: individual memory items can be reactivated simultaneously as a clustered occurrence or, alternatively, replayed sequentially as temporally separate instances. In the current study, participants learned associations between images that were embedded in a directed graph network and retained over a brief 8-minute consolidation period. During a subsequent cued recall session, participants retrieved the learned information while undergoing magnetoencephalographic (MEG) recording. Using a trained stimulus decoder, we found evidence for clustered reactivation of learned material. Reactivation strength of individual items during clustered reactivation decreased as a function of increasing graph distance, an ordering present solely for successful retrieval but not with retrieval failure. In line with previous research, we found evidence that sequential replay was dependent on retrieval performance and limited to low performers. The results provide further evidence for the existence of different performance-dependent retrieval mechanisms suggesting graded clustered reactivation as a plausible mechanism to search within abstract cognitive maps.

A remarkable capacity of the brain is its ability to autonomously reorganize memories during offline periods. Memory replay, a mechanism hypothesized to underlie biological offline learning, has inspired offline methods for reducing forgetting in artificial neural networks in continual learning settings. A memory-efficient and neurally-plausible method is generative replay, which achieves state of the art performance on continual learning benchmarks. However, unlike the brain, standard generative replay does not self-reorganize memories when trained offline on its own replay samples. We propose a novel architecture that augments generative replay with an adaptive, brain-like capacity to autonomously recover memories. We demonstrate this capacity of the architecture across several continual learning tasks and environments.

Deep neural networks have made tremendous gains in emulating human-like intelligence, and have been used increasingly as ways of understanding how the brain may solve the complex computational problems on which this relies. However, these still fall short of, and therefore fail to provide insight into how the brain supports strong forms of generalization of which humans are capable. One such case is out-of-distribution (OOD) generalization— successful performance on test examples that lie outside the distribution of the training set. Here, we identify properties of processing in the brain that may contribute to this ability. We describe a two-part algorithm that draws on specific features of neural computation to achieve OOD generalization, and provide a proof of concept by evaluating performance on two challenging cognitive tasks. First we draw on the fact that the mammalian brain represents metric spaces using grid cell code (e.g., in the entorhinal cortex): abstract representations of relational structure, organized in recurring motifs that cover the representational space. Second, we propose an attentional mechanism that operates over the grid cell code using Determinantal Point Process (DPP), that we call DPP attention (DPP-A) - a transformation that ensures maximum sparseness in the coverage of that space. We show that a loss function that combines standard task-optimized error with DPP-A can exploit the recurring motifs in the grid cell code, and can be integrated with common architectures to achieve strong OOD generalization performance on analogy and arithmetic tasks. This provides both an interpretation of how the grid cell code in the mammalian brain may contribute to generalization performance, and at the same time a potential means for improving such capabilities in artificial neural networks.

In addition to its critical role in encoding individual episodes, the hippocampus is capable of extracting regularities across experiences. This ability is central to category learning, and a growing literature indicates that the hippocampus indeed makes important contributions to this form of learning. Using a neural network model that mirrors the anatomy of the hippocampus, we investigated the mechanisms by which the hippocampus may support novel category learning. We simulated three category learning paradigms and evaluated the network’s ability to categorize and recognize specific exemplars in each. We found that the trisynaptic pathway within the hippocampus—connecting entorhinal cortex to dentate gyrus, CA3, and CA1—was critical for remembering exemplar-specific information, reflecting the rapid binding and pattern separation capabilities of this circuit. The monosynaptic pathway from entorhinal cortex to CA1, in contrast, specialized in detecting the regularities that define category structure across exemplars, supported by the use of distributed representations and a relatively slower learning rate. Together, the simulations provide an account of how the hippocampus and its constituent pathways support novel category learning.

The hippocampus is an archicortical structure, consisting of subfields with unique circuits. Understanding its microstructure, as proxied by these subfields, can improve our mechanistic understanding of learning and memory and has clinical potential for several neurological disorders. One prominent issue is how to parcellate, register, or retrieve homologous points between two hippocampi with grossly different morphologies. Here, we present a surface-based registration method that solves this issue in a contrast-agnostic, topology-preserving manner. Specifically, the entire hippocampus is first analytically unfolded, and then samples are registered in 2D unfolded space based on thickness, curvature, and gyrification. We demonstrate this method in seven 3D histology samples and show superior alignment with respect to subfields using this method over more conventional registration approaches.The methodological advancements described here are made easily accessible in the latest version of open source software HippUnfold. Code used in the development and testing of these methods, as well as preprocessed images, manual segmentations, and results, are openly available.