Predicting memory from the network structure of naturalistic events

The study examines how the web of interrelations among events in continuous, naturalistic experiences influences memory. Traditional memory experiments typically use isolated, randomized items, minimizing inter-item structure. In contrast, real-world experiences form coherent narratives with events linked semantically and causally across time. Prior work has emphasized event segmentation and boundary-related hippocampal and default mode network (DMN) responses, but less is known about how connections among temporally proximal and distal events shape encoding and retrieval. The authors hypothesize that events with stronger and more numerous semantic or causal connections within a narrative will be better remembered, and that this centrality will manifest in hippocampal boundary responses during encoding and in DMN activity and intersubject representational convergence during recall.

Past research on narratives and memory shows that event boundaries trigger hippocampal and cortical responses, and DMN patterns represent situation models reinstated during recall. Reading comprehension studies demonstrate that events or statements embedded in causal chains are remembered better, and perceived importance increases with causal connectivity. Inter-item semantic relationships in list learning can cue recall among related items. The authors leverage NLP advances (e.g., sentence embeddings) to quantify semantic similarity between complex event descriptions, complementing classic causality-based accounts. They position semantic and causal connectivity as potentially overlapping but distinct contributors to memory, integrating insights from event segmentation, DMN function in episodic recollection, and memory integration mechanisms.

Participants: 21 right-handed native English speakers (age 20–33; 12 females) were recruited; 6 were excluded for excessive motion, yielding N=15 for analyses. All provided informed consent. Stimuli and paradigm: Participants watched 10 short movies (3 animations, 7 live-action; mean length ~4.54 min) presented in two fMRI runs with fixed order. Each movie was prefaced by a title screen. Immediately afterward, participants performed free spoken recall in the scanner, describing the movie plots in their own words with no cues or guidance; cued recall followed but was not analyzed. Behavioral processing: Movies were segmented by an independent coder into 10–35 events (total 202). Three independent annotators provided fine-grained sub-event textual descriptions per event. Free recall was transcribed and segmented into utterances; only utterances recalling specific movie events were analyzed. Independent raters provided retrospective and rate-as-you-go importance ratings for each event. Narrative networks: For semantic networks, annotators’ event texts were converted into 512-d embeddings using Google’s Universal Sentence Encoder; cosine similarity defined edge weights between events within a movie. For causal networks, 12–13 independent coders per movie identified causally related event pairs; edge weights were the proportion of coders endorsing causality (direction ignored for primary analyses). Event centrality was defined as normalized degree (sum of edge weights) within each movie and z-scored. fMRI acquisition and preprocessing: 3T Siemens Prisma, multiband EPI (TR=1.5 s, TE=39 ms, voxel 2mm isotropic). Standard preprocessing included motion and fieldmap correction, coregistration, surface/volume mapping, smoothing (4 mm FWHM), high-pass filtering, nuisance regression for connectivity analyses, and z-scoring. Analyses: - Univariate GLM relating event-wise BOLD to semantic or causal centrality during watching (excluding first event of each movie) and recall; ROI comparisons (DMN parcels including angular gyrus and PMC; hippocampus; early visual cortex). - Intersubject pattern correlation (pISC) assessed event-specific pattern similarity across participants within cortical parcels during watching and recall; tested High (top 40%) vs Low (bottom 40%) centrality effects in ROIs (PMC, EVC) with randomization tests. - Representational similarity analysis (RSA) correlated cross-event neural similarity matrices during recall with semantic similarity matrices from annotations and recall transcripts across cortical parcels (FDR-corrected). - Hippocampal event boundary responses: time-locked BOLD to event onsets/offsets compared for High vs Low semantic centrality; logistic mixed-effects regression tested recall prediction; mediation analysis assessed whether hippocampal offset responses mediated centrality’s effect on recall. - Intersubject functional connectivity (ISFC): computed hippocampus–PMC/EVC connectivity within events ≥22.5 s and correlated ISFC with semantic centrality. - Low-level sensory controls: luminance, contrast, and audio amplitude tested for modulation by centrality. Replication: A preregistered online experiment (N=393; each participant watched one of 10 new short movies and provided written recall) replicated behavioral findings; narrative networks constructed similarly (single annotator per movie; 10 causality coders).

• Recall behavior: Participants recalled on average 9/10 movies (SD 1.2), with free recall lasting 32.4 min (SD 14.5). Within recalled movies, 77.6% of events were recalled (SD 11.2%), typically in original order (mean within-movie order correlation ρ=.97, SD .03); no primacy/recency effects (ANOVA F(2,18)=0.78, p=.47). - Semantic centrality and memory: Event recall probability correlated with semantic centrality r(202)=.20, p=.004. High (top 40%) vs Low (bottom 40%) semantic centrality events showed higher recall probability, t(14)=6.12, p<.001, Cohen’s d=1.58 (95% CI diff [.06, .12]). - Causal centrality and memory: Causal centrality correlated with recall probability r(202)=.29, p<.001; High vs Low causal centrality difference t(14)=8.23, p<.001, d=2.12 (95% CI [.10, .17]). Mixed-effects logistic regression indicated unique contributions of semantic centrality (β=.17, SE=.05, χ²(1)=12.24, p<.001) and causal centrality (β=.38, SE=.05, χ²(1)=55.04, p<.001). Semantic and causal centrality correlated r(202)=.28, p<.001. - Replication (online): Both semantic (β=.17, SE=.03, χ²(1)=48.52, p<.001) and causal (β=.44, SE=.03, χ²(1)=255.67, p<.001) centrality uniquely predicted recall; no serial position effects (F(2,18)=.85, p=.44). - fMRI univariate: During recall, higher semantic centrality associated with stronger activation in DMN regions including angular gyrus and PMC; hippocampus also showed higher activation for high centrality events t(14)=2.71, p=.017, d=.70. During movie watching, at liberal threshold (uncorrected p<.001), centrality scaled activation in visual/auditory association cortices and precuneus; low-level sensory features were not significantly modulated by centrality (all χ²(1)<1.94, ps>.16). Causal centrality showed similar DMN recall effects. - Event-specific intersubject pattern similarity (pISC): Robust pISC during recall in DMN, especially PMC. During recall, PMC pISC was higher for High vs Low semantic centrality (diff=.019; two-tailed randomization p=.037). EVC showed the opposite pattern (diff=−.013; p=.012). During movie watching, centrality-related pISC differences were not significant in these ROIs. - RSA: Cross-event neural similarity during recall correlated with semantic similarity from annotations and more strongly with recall transcripts, predominantly in DMN parcels (notably PMC) after FDR correction. - Hippocampal boundary responses and mediation: Hippocampal BOLD responses following event offset were higher for High vs Low semantic centrality; onset responses did not differ. Stronger offset responses predicted recall (β=.26, SE=.10, χ²(1)=6.37, p=.012). Mediation analysis showed hippocampal offset responses partially mediated the effect of semantic centrality on recall (average causal mediation effect=.001, p=.016; direct effect remained significant β=.20, SE=.05, χ²(1)=13.91, p<.001). - Hippocampus–cortex ISFC: Hippocampus–PMC ISFC increased with semantic centrality (r(26)=.49, p=.01); hippocampus–EVC ISFC did not (r(26)=.01, p=.95), and the difference in correlations was significant (95% CI of difference [.05, .87]). - Importance ratings: Semantic centrality correlated with perceived importance (r(202)=.22, p=.002).

Findings show that the inter-event structure of narratives, quantified as semantic and causal centrality within event networks, predicts both what is remembered and the neural signatures during encoding and recall. Higher centrality likely benefits recall by increasing the number/strength of cues from other events and by preserving narrative coherence when retelling. During encoding, high centrality events may be repeatedly reactivated via shared components, strengthening integration and leading to stronger hippocampal boundary responses and greater hippocampus–PMC coupling. During recall, DMN regions (PMC and angular gyrus) showed greater activation and more convergent event-specific representations across individuals for high centrality events, consistent with richer, more accurate episodic reconstruction. Opposite convergence in EVC during recall suggests shifts between internal and external processing modes across centrality levels. Although semantic and causal centrality overlapped, each contributed uniquely to memory; weaker multivoxel effects for causality may reflect sparser, adjacent, and more idiosyncratic causal links in film stimuli and differences in measurement (text embeddings vs. human judgments). The results underscore that narrative network position shapes memory reliability and accessibility, extending beyond classic serial position effects from random-item lists and highlighting the role of structural coherence in naturalistic memory.

Transforming movies into semantic and causal event networks reveals that events central in these networks are better remembered and show distinct neural signatures: stronger hippocampal boundary responses and hippocampus–PMC coupling during encoding, and heightened DMN activation and intersubject representational convergence during recall. This demonstrates that inter-event structure shapes the encoding, storage, and retrieval of complex experiences. The approach suggests practical applications for memory interventions and education by leveraging network structure to enhance memory. Future work should orthogonalize different inter-event relations (semantic, causal, emotional), examine evolving centrality during unfolding narratives, probe generalization to real-life events without narrative design constraints, and detail mechanisms by which schemas and predictive cues inform perceived centrality and neural dynamics.

• Sample size for fMRI analyses was modest (N=15 after exclusions), which may limit generalizability. - ISFC analyses included only 26 sufficiently long events, reducing statistical power; replication with larger datasets is needed. - Causal networks were relatively sparse and mostly linked adjacent events; causality judgments were more idiosyncratic across coders than semantic similarity, potentially weakening multivoxel effects. - Semantic and causal centrality were measured differently (sentence embeddings vs. human judgments), complicating direct comparisons; the study did not fully dissociate their unique mechanisms. - Intersubject convergence effects in EVC during recall were very small overall, and sensory cortex recall signals are weak without shared visual input. - Stimuli were fictional movies with possible design pressures that reduce interpretive variability for central events; generalization to everyday, non-scripted events remains to be tested. - Although low-level sensory features did not vary with centrality on measured metrics, unmeasured higher-level perceptual factors could contribute. - First events were excluded from certain analyses to avoid boundary artifacts, which may bias estimates of early-narrative effects.