Dynamic patterns of functional connectivity in the human brain underlie individual memory formation

Episodic memory formation integrates information about an experience’s content and spatiotemporal context into a unique memory trace. Because different experiences recruit different sets of brain regions and temporal dynamics, encoding distinct episodes should involve distinct and dynamically changing patterns of communication between regions. Prior work has linked large-scale connectivity to memory but has struggled to interpret rapid, event-specific modulations that vary across trials. A key challenge is determining whether transient connectivity changes observed during individual events are meaningful, and whether similar patterns are reinstated during retrieval of the same memory. The authors hypothesize that sub-second, event-specific changes in functional connectivity support encoding of individual word-pair associations and that these dynamic patterns are reinstated during successful retrieval. Using intracranial EEG, they seek to identify reliable functional connections and test whether moment-to-moment coupling patterns during encoding recur during retrieval, beyond generic task-related processes.

Previous research has characterized large-scale functional connectivity and its relation to memory using neuroimaging and electrophysiology. Studies with scalp and intracranial EEG have linked dynamic sub-second changes, particularly in theta and high-gamma bands, to memory formation and retrieval, often by aggregating across many trials. Work on oscillatory phase reinstatement during retrieval suggests that coordinated phase relations might underlie communication between regions and could be reinstated with memory. However, interpreting heterogeneous, event-specific connectivity modulations remains difficult, as many approaches average over events and obscure moment-to-moment variability. This study builds on findings of stable, long-term functional networks with consistent time delays and coherence structure, aiming to resolve whether event-specific dynamic connectivity patterns during encoding are reinstated during successful retrieval and whether such reinstatement is distinct from reinstatement of local spectral power.

Participants: Twenty adults with drug-resistant epilepsy (8 females; 33.40 ± 2.30 years; IQ > 70) undergoing clinical intracranial monitoring participated. Subdural electrodes (grids/strips; inter-contact spacing 5–10 mm) covered widespread cortical areas, with consistent anterior temporal coverage. Electrode locations were obtained by co-registering post-op CT with pre-op MRI. Inclusion criteria included no prior resection, at least 10 accurate trials across sessions, and artifact-free data. IRB approval and informed consent were obtained. Task: A paired-associates verbal memory task was used. Each session comprised 25 lists of 6 word pairs (common nouns from a 300-word pool), presented sequentially. Encoding: fixation (+) for 250 ms, ISI 500–750 ms, word pair for 4000 ms, ISI 1000 ms. A 20 s arithmetic distractor followed each list. Retrieval: cue (????) 250–300 ms, ISI 500–750 ms, cue word for 4000 ms; participants vocalized the associate at any time during retrieval. Trials with response latency > 4000 ms were excluded. Responses were labeled correct, intrusion, or pass; intrusion and pass were considered incorrect. For no-vocalization trials, a response time was sampled from that session’s correct RT distribution to align trials (retrieval time = 0 s at vocalization onset for correctly timed responses). iEEG acquisition and preprocessing: Signals were recorded with Nihon Kohden or Blackrock systems at 1000–2000 Hz and resampled to 1000 Hz. Electrodes with abnormal amplitudes/variance (>3 SD) or excessive line noise were rejected. Signals underwent local detrending, regression-based 60/120 Hz line noise removal, low-pass at 150 Hz, and bipolar re-referencing between adjacent contacts (defining electrode locations at midpoints). In total, 1536 bipolar electrodes were retained (31–132 per participant; mean 76.80 ± 5.25). Epochs extracted: encoding −2000 to +5000 ms from study onset; retrieval −5000 to +2000 ms relative to vocalization (or assigned time), with 1000 ms buffers removed after processing. Spectral power: Time–frequency power (3–150 Hz) was computed via complex Morlet wavelets (40 logarithmically spaced frequencies; wavelet number 6), squared and log-transformed. Power was baseline-normalized per electrode and frequency using −700 to −500 ms pre-stimulus. Identifying functional connections: To estimate directed functional connectivity agnostic to task timing, 20 random 30 s blocks (∼1/6 of session) were sampled per session. Broadband (1–150 Hz) absolute cross-correlations between electrode pairs were computed in overlapping 1000 ms windows (80% overlap; step 200 ms) across temporal offsets τ = −200 to +200 ms (1 ms steps). The coupling function W(τ) was the z-scored cross-correlation over τ. Averaging across blocks yielded a robust coupling profile per pair; the maximum Wmax at delay τmax indexed consistent, time-locked coupling. Peak sharpness was quantified via a coincidence index (CI; normalized full width at half-maximum area of the dominant peak). Thresholds for Wmax and CI were defined per participant by fitting two-component Gaussian mixtures and selecting the knee point between component means; pairs with τmax = 0 were excluded to mitigate volume conduction. Pairs had to meet thresholds across multiple sessions (hours–days apart) to ensure stability. Spectral coherence for identified pairs was computed over the task to estimate characteristic frequencies. Time-varying coupling during task: For each identified pair, sliding-window coupling W at τmax was computed in 1000 ms windows (80% overlap) throughout encoding and retrieval. Coupling time series were baseline-normalized (Wz) per pair using across-trial mean and SD from −1000 to −500 ms before stimulus onset. Pairs were categorized as increasing or decreasing based on average Wz during encoding (0–4000 ms). The magnitude of coupling change was assessed as |Wz| over time, averaged per pair and then across pairs/trials (separately for correct/incorrect). Neural reinstatement analyses: For each time point i (encoding) and j (retrieval), feature vectors were constructed: (a) coupling reinstatement used Wz values from all functionally connected pairs at their τmax; (b) spectral power reinstatement used baseline-normalized power across all electrodes in five bands (theta 3.5–8, alpha 8–12, beta 13–25, low gamma 30–58, high gamma 62–100 Hz), concatenated across electrodes and bands. Encoding–retrieval reinstatement S(i,j) was quantified as cosine similarity between feature vectors for every (i,j) pair, yielding a temporal reinstatement map per trial. Maps were averaged within correct and incorrect trials per participant. Temporal regions of interest (tROIs) were defined separately for coupling and power by cluster-based permutation tests identifying (encoding, retrieval) time pairs (pre-vocalization) with significantly greater similarity for correct vs incorrect. Spectral power vs coupling separability: A leave-one-out procedure estimated each electrode’s relative contribution to coupling and power reinstatement within the respective tROIs by recomputing correct–incorrect reinstatement differences after excluding an electrode (for power) or electrode pair (for coupling), then averaging pairwise contributions per electrode for coupling. Contributions were normalized to [0,1] within participant. Correlations between electrodes’ contributions to coupling vs power reinstatement were computed per participant, with attenuation correction estimated via resampling-based reliability. Median splits of electrodes by contribution (to coupling or power) examined whether grouping by one measure affected the other; two-way ANOVAs tested interactions. Statistics: Participant-level inference used paired t-tests, repeated-measures ANOVAs, and cluster-based permutation tests (1000 permutations; two-tailed; alpha = 0.05) to define significant clusters and tROIs and to compare correct vs incorrect reinstatement and |Wz| time courses.

- Behavioral: Participants correctly recalled 31.22 ± 4.91% of word pairs; median reaction time 2103 ± 105 ms. - Identification of functional connections: On average, 371.49 ± 33.93 strongly connected electrode pairs per session (11.60% ± 0.62% of possible). Requiring stability across sessions reduced this to 184.60 ± 26.23 per participant (6.35% ± 0.65%). Spectral coherence of connected pairs peaked around 6–10 Hz (peak ~7 Hz across participants). - Dynamic coupling during task: The magnitude of coupling change |Wz| increased from baseline during encoding and peaked around −1000 to −500 ms relative to retrieval vocalization (retrieval time = 0 s). Encoding |Wz| was significantly greater for correct vs incorrect trials from 400 to 3200 ms after study onset (cluster-based Pcorrected < 0.05; mean difference t(19) = 3.18, p = 0.0049, Cohen’s d = 0.71, 95% CI [0.21, 1.20]). During retrieval, |Wz| increased before vocalization for both correct and incorrect trials, without a significant correct–incorrect difference. - Reinstatement of dynamic coupling: Encoding–retrieval similarity of coupling patterns was significantly greater for correct vs incorrect trials, peaking at 1360 ± 200 ms post-study onset and −550 ± 170 ms relative to vocalization (cluster-based Pcorrected < 0.05; within tROI: t(19) = 3.60, p = 0.0019, d = 1.03, 95% CI [0.48, 1.47]). Reinstatement was specific to item identity: original correct trials > shuffled correct retrieval labels (t(19) = 3.55, p = 0.0021, d = 0.79, 95% CI [0.28, 1.29]) and > adjacent correct retrieval trials (t(19) = 3.91, p = 0.00093, d = 0.87, 95% CI [0.35, 1.39]). - Separability from spectral power reinstatement: Power reinstatement also showed correct > incorrect before vocalization (tROI peaking ~1050 ± 340 ms after study onset and −1000 ± 330 ms before vocalization). However, electrodes’ contributions to coupling vs power reinstatement were not systematically correlated (average Fisher’s r = 0.021, t(19) = 0.83, p = 0.42; attenuation-corrected r = 0.051, t(19) = 0.67, p = 0.51). Median split by coupling contribution increased coupling reinstatement but did not significantly change power reinstatement (t(19) = 2.06, p = 0.053); significant interaction between electrode group and reinstatement measure: F(1,19) = 37.52, p = 0.0000069, ηp² = 0.66. Conversely, median split by power contribution increased power reinstatement but not coupling reinstatement (t(19) = 1.45, p = 0.16); interaction: F(1,19) = 18.21, p = 0.00042, ηp² = 0.49. Both high and low groups still showed reliable reinstatement of the other measure within its tROI. - Spatial distribution: Electrode pairs showing increases and decreases in coupling were distributed across overlapping cortical regions, without evidence for specific hubs preferentially contributing to reinstatement.

The study demonstrates that sub-second, event-specific patterns of functional connectivity across cortical regions are a key component of episodic memory. During encoding, coupling strengths between reliably connected electrode pairs change dynamically; successful retrieval involves reinstatement of the specific pattern of these dynamic changes that occurred during encoding of the same item. This reinstatement is stronger in correct trials and remains item-specific relative to shuffled or adjacent correct comparisons, indicating it is not merely a generic retrieval state. Importantly, coupling reinstatement is complementary to but separable from reinstatement of local spectral power: electrodes contributing most to coupling reinstatement do not necessarily contribute to power reinstatement, and vice versa. These findings extend the concept of neural reinstatement beyond local activity to include distributed, time-resolved network interactions, suggesting that memory retrieval recapitulates not only local representations but also the temporal coordination among regions that supported encoding. The work addresses a central challenge in dynamic connectivity research by linking transient, heterogeneous connectivity fluctuations to specific mnemonic content and behavior.

This work provides direct evidence that successful episodic memory formation and retrieval rely on specific, sub-second patterns of functional connectivity that are reinstated at retrieval and are distinct from local spectral power reinstatement. By identifying stable functional connections via consistent time delays and tracking their moment-to-moment coupling during a paired-associates task, the study shows that item-specific connectivity patterns during encoding are reinstated prior to correct recall. These results emphasize that memory involves coordinated, dynamic interactions across distributed cortical networks in addition to local activations. Future research could probe causal roles of specific connections, explore how such dynamic patterns interact with hippocampal dynamics and other frequency bands, and test whether targeted neuromodulation can enhance memory by selectively manipulating event-specific connections.

- Participant cohort: All participants were patients with drug-resistant epilepsy undergoing clinical monitoring; electrode placement was determined by clinical needs, leading to non-uniform cortical coverage and potential limits on generalizability to the healthy population. - Recording coverage and modality: Analyses focused on subdural (cortical) electrodes; deep structures (e.g., hippocampus) were not systematically sampled in this dataset, potentially missing contributions from medial temporal lobe circuits. - Connectivity inference: Functional connections were identified via cross-correlation with consistent time delays and zero-lag pairs were excluded to mitigate volume conduction. While informative about time-locked coupling, these measures do not establish causal directionality or the informational content of interactions. - Task specificity and trial counts: The study used a verbal paired-associates task with response-times constraints and required at least 10 accurate trials per participant; findings may not generalize to other memory tasks or stimulus types. - Heterogeneity and sparsity: Only a sparse subset of all possible electrode pairs met stringent stability thresholds across sessions, which may bias analyses toward stronger or more stable connections.