Hippocampal systems for event encoding and sequencing during ongoing narrative comprehension

The study addresses how the human brain encodes discrete narrative events and sequences them into coherent structures during continuous, naturalistic comprehension. Narratives consist of temporally discrete but causally and semantically connected events; successful comprehension requires capturing detailed event content and inferring temporal and causal relationships among events. Prior neuroimaging work implicates the hippocampus and default mode network (DMN), especially the posterior medial cortex (PMC), in event encoding at boundaries, and the hippocampus and medial prefrontal cortex (mPFC/vmPFC) in integrating information across episodes. Building on computational principles that memory encoding is optimal at event boundaries and that sequencing occurs when current events require past context, the authors hypothesized two distinct hippocampus-centered processes: 1) hippocampus–PMC coupling supporting content encoding following event boundaries, and 2) hippocampus–vmPFC interactions supporting event sequencing during moments requiring integration of current and past information. The purpose is to identify when and how these systems operate during ongoing narrative processing without relying solely on experimenter-imposed structure, and to quantify individual differences in content and sequence memory from free recall. Understanding these mechanisms is important for elucidating real-world memory construction from fragmented experiences.

Neuroimaging with naturalistic stimuli (movies, stories) shows hippocampus and DMN engagement at event boundaries, with PMC representing event structure and exhibiting similar patterns during viewing and recall; hippocampal post-boundary activity predicts later memory. Beyond individual events, integration across episodes is supported by hippocampus and mPFC; lesions in this circuit impair integration of temporally scattered information, and hippocampus–mPFC synchrony modulates schema-consistent vs. novel information encoding. Computational models suggest selective encoding at boundaries and dynamic sequencing when resolving uncertainty by integrating past context; fMRI studies show reactivation of semantically related past events in the hippocampus to support integration. These lines of work motivate testing boundary-based encoding via hippocampus–PMC coupling and sequencing via hippocampus–vmPFC connectivity during ongoing narratives.

Participants: 71 recruited (26 females; mean age 22.78 ± 2.28 years); 6 excluded (1 global artifact; 5 excessive motion: framewise displacement > 0.5 mm for >5% of images), yielding N = 65 for analysis. Informed consent obtained; ethics approved by Sungkyunkwan University IRB; none had seen the stimulus prior. Stimulus: 10-min audiovisual animated movie, “Mr. Bean: The Animated Series, Art Thief” (S2E13, 2003), comprising 17 events (~36 s each) defined primarily by director’s cuts. Events were temporally scrambled in an identical pseudorandom order across participants to increase sequencing demands; included 12 event boundaries with pairs of related events occurring before/after their counterpart in the scrambled order. A 30 s nature video preceded the movie as a buffer. Procedure: fMRI session had four functional runs and one anatomical run. Analyses focused on the first run: scrambled movie viewing (10 min), followed by free recall where participants recounted the inferred original story in chronological order with unlimited time (mean recall duration 3.59 min, SD 2.03). Participants then viewed the intact movie, then scrambled movie again, followed by second recall (not analyzed here), and finally a resting-state run. A practice session with a different scrambled cartoon ensured task understanding. Narrative memory measures: Four annotators created detailed annotations of the original movie at 2 s resolution; free recall was transcribed at 5 s resolution. Text preprocessing in Korean (KoNLPy): tokenize, POS-tag, extract nouns/verbs; apply a custom dictionary (67 word mappings) to align synonyms (e.g., wrench/spanner). Sentences aggregated by window size and converted to vectors via bag-of-words (scikit-learn) for topic modeling (LDA). Topic model trained on movie annotations; both annotations and recall mapped to topic vectors; cosine similarity computed to form movie–recall similarity matrices at the sentence level. Two scores derived: (1) Ordering score (So): convert similarity matrix to binary matches with a threshold and compute Spearman rank correlation between recall order and original chronological order of matched indices; (2) Content score (Sc): compute recall content distribution (average topic similarities across recalled elements for each movie time segment) and movie content distribution (from movie–movie similarity); Sc = 1/(1 + DKL) between these distributions, range 0–1. Human validation by two independent raters who segmented events (46 and 25 events) and matched recall to annotation; hyperparameters optimized (80 topics; window size 0 for both; threshold 0.3) to maximize correlations with human-rated scores; robustness assessed across parameter ranges. fMRI acquisition: 3 T Siemens Prisma, 64-channel head coil. T2*-weighted EPI: voxel 3 mm isotropic, TR 1000 ms, TE 30 ms, FOV 240 × 240 mm, 48 slices whole-brain, flip angle 90°, multiband factor 3. T1-weighted MPRAGE: 1 mm isotropic, TR 2200 ms, TE 2.44 ms, FOV 256 × 256 mm, 256 slices, flip angle 8°. Preprocessing: fMRIPrep pipeline: anatomical INU correction, skull-stripping, segmentation, surface reconstruction; motion correction; registration to MNI152; denoising using motion parameters and derivatives, global signal, framewise displacement, physiological regressors (aCompCor). Spatial smoothing (FWHM 5 mm) and intensity normalization in AFNI. FC-based predictive modeling: Connectivity computed as Pearson correlations (Fisher z-transformed) between ROI time courses. Hippocampus from Brainnetome atlas (left/right anterior/posterior subregions averaged to single time course); cortical ROIs from Schaefer atlas with multiple parcellations (main analysis: 200 parcels). Periods for connectivity estimation: (1) Expected sequencing moments (4 s per pair at event beginning/end depending on original order; total 64 time points), (2) Post-event boundary moments defined by elevated hippocampal activity 4–7 s after boundaries (4 s per event; 68 time points), and (3) All movie time points (610 TRs including blank screen). Hippocampo–cortical model: hippocampus seed to each cortical ROI (200 edges). Cortico–cortical model: all inter-cortical edges excluding hippocampus (19,900 edges), with control matching the number of edges to the hippocampal model. Leave-one-subject-out cross-validation: within training folds, select edges correlated with target memory score (p < 0.05); sum Fisher z-values of positively and negatively correlated edges as predictors; fit linear regression to predict the score; evaluate Pearson correlation between predicted and actual in left-out subject. Significance assessed by permutation tests (1000 iterations shuffling scores). Edge selection consistency assessed via one-tailed binomial tests with FDR correction (q < 0.001). LLM-derived sequencing moments: Detailed 2 s segment annotations fed to BERT (bert-base-uncased) to compute next sentence prediction (NSP) likelihood between each segment and all preceding segments. For each segment, take the maximum NSP likelihood over past segments as the segment’s NSP score. Control for semantic similarity via Universal Sentence Encoder (USE) cosine similarity by regressing it out from NSP to yield a narrative coherence index, convolved with a canonical HRF. Validate coherence against human-rated moment-by-moment narrative comprehension (from prior study with 20 participants) via Pearson correlation and phase-randomization permutation tests (n = 1000). Define LLM-generated sequencing moments as top-N time points ranked by coherence (varying N from 20 to 80; main results use top 50). Recompute hippocampo–cortical FC over these moments and apply same predictive modeling pipeline. Comparisons also made using raw NSP-only and semantic similarity-only moments. Data and code: fMRI data available on OpenNeuro (ds005215); behavioral/code on GitHub and Zenodo; models available on Hugging Face.

• Two independent memory metrics from topic modeling of movie annotations vs. free recall captured distinct aspects of narrative comprehension: content scores (semantic overlap) and ordering scores (temporal sequencing). They were not significantly correlated (r = 0.173, p = 0.166).
• Hippocampo–cortical FC models predicted both metrics, whereas cortico–cortical models (excluding hippocampus) did not: • Ordering scores predicted during sequencing moments: cross-validated r = 0.314, p = 0.023 (one-tailed). • Content scores predicted during post-event boundaries: r = 0.299, p = 0.034; using all time points yielded marginal performance r = 0.247, p = 0.067. • Cortico–cortical models failed to significantly predict either score, including control analyses matching the number of edges.
• Critical edges consistently selected across folds and parcellations: • Hippocampus–vmPFC FC exclusively predicted ordering scores during sequencing moments (selected in 100% of CV folds for ordering; 0% for content; χ²(1, N = 65) = 130.0, p < 0.001). • Hippocampus–PMC FC predominantly predicted content scores during post-boundary moments (selected in 87.6% of CV folds for content; 0% for ordering; χ²(1, N = 65) = 101.5, p < 0.001).
• Correlation patterns supported dissociable systems: • Hippocampus–vmPFC FC negatively correlated with ordering scores during sequencing moments (r = −0.369, p = 0.002) and did not relate to content (r = 0.036, p = 0.77). Steiger’s Z showed stronger correlation with ordering than content (z(62) = 2.614, p = 0.008). • Hippocampus–PMC FC positively correlated with content scores during post-boundaries (r = 0.258, p = 0.037) and did not relate to ordering (r = 0.029, p = 0.812); difference between correlations was not significant (z(62) = 1.436, p = 0.15).
• Model robustness: • Inclusion of more cortical ROIs reduced predictive performance when hippocampus was included (ordering: r = −0.921, p < 0.001; content: r = −0.902, p < 0.001), suggesting importance of functionally relevant edges; cortico–cortical models underperformed regardless of ROI number. • Hippocampus-seeded models outperformed most other seed regions across parcellations (permutation-based rank analyses; all ps < 0.05), with the hippocampus achieving highest accuracy for predicting ordering.
• LLM-derived narrative coherence identified data-driven sequencing moments: • LLM coherence correlated with human-rated moment-by-moment narrative comprehension (r = 0.259, p = 0.016); raw NSP likelihood (r = 0.187, p = 0.06) and semantic similarity alone (r = 0.08, p = 0.212) did not. • Hippocampo–cortical model trained on top 50 LLM-coherence moments predicted ordering scores (r = 0.305, p = 0.034). Models using high raw NSP likelihood or high semantic similarity alone failed (NSP: r = −0.026, p = 0.51; semantic: r = 0.036, p = 0.42). • Performance declined as the number of LLM-selected moments increased beyond top moments, indicating sequencing occurs at specific high-coherence times. • Hippocampus–vmPFC FC remained a significant, negatively correlated predictor of ordering during LLM moments (r = −0.301, p = 0.014), with no relation to content (r = −0.104, p = 0.409). Approximately 15% overlap existed between pre-defined and LLM-identified sequencing moments; pre-defined moments showed higher coherence than other segments (t(608) = 2.91, p = 0.003). Overall, results reveal two distinct hippocampus-centered systems: hippocampus–PMC supports content encoding at event boundaries, and hippocampus–vmPFC supports integration and sequencing during moments of high narrative coherence.

The findings demonstrate that ongoing narrative comprehension relies on distinct hippocampus-centered systems operating at specific times. Hippocampus–PMC coupling at event boundaries supports detailed content encoding, consistent with DMN engagement in representing event structure and boundary-triggered memory formation. Hippocampus–vmPFC interactions during sequencing moments support integrating current events with prior context and reconstructing temporal order; notably, stronger desynchronization (negative FC correlation with performance) aligns with integrating schema-consistent information rather than encoding novel, schema-incongruent events. FC-based predictive modeling across time windows and parcellations isolates these key features, extending evidence from lesion, electrophysiology, and neuroimaging studies that implicate hippocampus and mPFC in sequencing and integration across episodes. A data-driven LLM approach identifies moments of high narrative coherence beyond pre-defined structural points, correlating with human comprehension and pinpointing when the brain likely sequences related events. This method validates the role of hippocampus–vmPFC coupling in sequencing and offers a tool to analyze naturalistic datasets without experimental manipulation. Collectively, the results illuminate dynamic memory processes by which the brain encodes complete events and integrates them into coherent narratives during naturalistic experiences.

This work provides converging behavioral, computational, and neuroimaging evidence for two dissociable hippocampal systems in narrative memory: hippocampus–PMC coupling supports content encoding at event boundaries, and hippocampus–vmPFC interactions support sequencing through integration of past and present events during moments of high narrative coherence. The authors introduce validated topic-model-based metrics to quantify content and order memory from free recall and show that hippocampo–cortical FC patterns predict individual differences in each process. A novel LLM-based coherence index reveals additional sequencing moments and aligns with human comprehension, underscoring the importance of narrative coherence in real-time integration. Future directions include directly probing sequencing operations (retrieval, reactivation, integration) during viewing, varying scrambling schemes, and testing longer, more complex narratives to assess generalizability. Combining neuroimaging with in-task behavioral measures or computational models may further dissociate component processes and refine neural mechanistic accounts of real-world memory construction.

• Sequencing performance was inferred rather than directly measured during viewing; no explicit sequencing task was included to preserve naturalism.
• Component cognitive operations (retrieval, reactivation, integration) underlying sequencing were not independently dissociated.
• Fixed scrambling order and a limited number of narrative events may constrain generalizability; narrative discontinuities in natural experiences are typically subtler.
• While LLM-derived coherence offers a data-driven approach, further validation across diverse narrative contexts is needed.