Hippocampal neurons code individual episodic memories in humans

Episodic memory enables reinstatement of the what, where and when of past experiences and is thought to depend on reinstating neural activity present at encoding. The hippocampus is central to binding multimodal information, but it is controversial whether hippocampal neurons code specific invariant elements (concept neurons) or entire episodes via conjunctive codes. Concept neurons respond to specific items independent of context, suggesting episodes might be represented by simultaneous activity of multiple such neurons or expanded selectivity. Alternatively, the indexing theory proposes that sparse hippocampal units bind together the elements of an individual episode and act as pointers to cortical modules. The study tests whether single hippocampal neurons code entire episodes rather than individual elements by examining reinstatement of firing at retrieval for specific encoded episodes.

Prior work has identified concept neurons in the human medial temporal lobe that respond invariantly to specific persons, places or objects across contexts and can increase firing during retrieval if their preferred image is part of the memory. Computational and theoretical accounts, including hippocampal indexing theory, propose that hippocampal neurons may act as indices to bind episode elements. Functional MRI studies have shown item-specific activity reinstatement in hippocampus, but fMRI cannot disambiguate whether reinstatement arises from concept neurons or indexing neurons. Human single-unit studies have also reported neurons responsive to event boundaries, novelty, and retrieval confidence, often across many episodes, and time cells that fire at specific moments within sequences. These literatures set the stage to test for neurons that specifically and conjunctively code unique episodes, independent of concept or time coding.

Participants were epilepsy patients implanted with Behnke-Fried depth electrodes with microwire bundles targeting the hippocampus and, in some sessions, parahippocampus. Two experiments were conducted. Experiment 1 recorded 585 hippocampal neurons from 16 participants (7 female; mean age 36.125 years). Encoding: participants created a vivid mental story linking an animal cue to two associate images (faces and/or places). Two seconds after cue onset, associate images appeared, and participants rated plausibility. A distractor task (odd/even judgments; 15 trials; ~22–225 s) followed. Retrieval: animal cues were presented, and participants indicated and then selected associated images from arrays. Experiment 2 recorded 216 hippocampal neurons from 14 participants (7 female; mean age 33.857 years). It used one cue–one associate pairs (animal/face/place) with 4-alternative forced-choice retrieval. After memory testing, Experiment 2 included a visual tuning task in which all prior images were presented six times (1 s each; 500–550 ms ISI) for category judgments, enabling identification of visually tuned concept neurons. Electrophysiology: Signals were acquired at 32 kHz or 32.768 kHz, filtered (0.1–9000 Hz analog; 300–3000 Hz digital for spikes). Spike detection extracted 2 ms waveforms around peaks, aligned and feature-extracted via wavelet decomposition; clustering used superparamagnetic clustering (wave_clus), with manual curation. Units were classified as single-unit or multi-unit using inter-spike interval and waveform variability criteria. ESN identification (rate code): For each neuron, firing rates per episode were computed. Encoding window: from associate image onset (2 s after cue) to episode end. Retrieval window: from cue onset to response about number remembered. Firing rates were z-scored within encoding and retrieval separately. Only later-remembered (hit) episodes were included for ESN analysis; later-forgotten episodes for miss-ESN analysis. Episode-specific reinstatement was defined as the elementwise product of z-encoding and z-retrieval rates. A neuron-specific threshold was derived by shuffling encoding–retrieval episode order 10,000 times and taking the 99th percentile. ESN criteria: (1) empirical reinstatement exceeded threshold for at least one episode; (2) both encoding and retrieval z-scores ≥ 1.645 for that episode to prevent dominance by one phase. In Experiment 1, neurons showing significant cue-evoked responses to the animal during the first second of encoding (concept neuron proxy) were excluded. In Experiment 2, episodes where a neuron showed significant visual tuning (from the tuning task) to either cue or associate were excluded. Robustness checks used alternative reinstatement measures (sum E+R; higher z-threshold 2.6; normalized product (E×R)/|E−R|). Temporal ESNs (tESNs): For each neuron, instantaneous firing rates were computed by convolving spikes with a Gaussian kernel (σ=100 ms) from 6 s before to 1 s after response at encoding and retrieval (first/last second excluded). Cross-correlation between encoding and retrieval instantaneous rate traces (±2.5 s lag) yielded a maximum value as the reinstatement metric per episode. A 99th percentile threshold was derived from 1,000 shuffles of episode order; neurons exceeding the threshold for at least one episode were marked as tESNs. For Experiment 2, trials with visual tuning were excluded. Validity was assessed with random and circularly shuffled surrogate spike times. Concept neuron detection: Significant visual responses were identified by comparing 19 overlapping 100 ms post-stimulus bins within 1,000 ms after onset against a −500–0 ms baseline using Mann–Whitney U-tests with Simes correction. Thresholds included a standard P<0.0005 and a liberal P=0.05 (uncorrected across images) to test robustness. Time cells (TCs): Encoding blocks were divided into 40 normalized bins or analyzed in non-normalized time; Kruskal–Wallis tests across bins and circular shift permutation tests assessed time specificity. Electrode localization used coregistered pre-op MRI and post-op MRI/CT to assign microwires to hippocampus or parahippocampus. Statistical inference used permutation tests (typically 10,000 draws) with group-level set-based comparisons of ESN counts against shuffled data. Behavioral performance was ~68% correct in Experiment 1 and ~66% in Experiment 2, above chance.

• Significant numbers of hippocampal episode-specific neurons (ESNs) reinstated firing for specific remembered episodes: • Experiment 1: 136/585 neurons (23.25%), P<0.001 (permutation test). 86.03% (117/136) coded a single episode. • Experiment 2 (excluding visually tuned responses): 38/216 neurons (17.59%), P=0.0053. 89.47% (34/38) coded a single episode. • Robustness: Alternative reinstatement measures yielded comparable results: E+R criterion (125 ESNs; P<0.001), stricter z≥2.6 (29 ESNs; P<0.001), normalized product (53 ESNs; P<0.001).
• ESNs were not driven by visual content or cue responses: • Experiment 1 excluded neurons responsive to the animal cue at encoding. • Experiment 2 visual tuning task identified concept neurons; even with a liberal uncorrected P=0.05 threshold inflating concept neuron counts (58→155/216), ESN count remained similar (36/216, 16.67%; P=0.0025). Only 4–6 potential ESNs were excluded due to tuning.
• Regional specificity: No significant ESNs in parahippocampus (Exp. 1: 15/104, P=0.5396; Exp. 2: 3/25, P=0.1199), consistent with hippocampal specificity (caution due to limited sampling).
• Temporal code reinstatement: • Significant tESNs were found: Experiment 1: 100/585 (17.09%), P=0.016; Experiment 2: 40/216 (18.52%), P<0.01 (excluding visually tuned trials). • Overlap between rate-code ESNs and tESNs at the episode level was significant: in Exp. 1, 20.25% of ESN-reinstated episodes also reinstated temporally and 25.81% of tESN-reinstated episodes also reinstated by rate (both P<0.001); Exp. 2: 26.19% and 24.44% overlaps (both P<0.001). • Control analyses with random spike times and circular shuffles did not yield significant tESN counts (≤5%).
• ESNs linked to memory success: No significant ESNs when restricting to later-forgotten episodes (miss-ESNs: 15/585, 2.56%; P=0.4229). After equalizing event numbers via bootstrapping, hit vs miss ESN counts did not differ (P=0.7032).
• ESNs are distinct from time cells: Few neurons met TC criteria (normalized: 12/585; non-normalized: 10/585), below chance (P>0.9), with no significant overlap between TCs and ESNs (P>0.3).
• Physiological properties: ESNs exhibited wider spike waveforms than other units (Experiment 1 P=0.0563; combined experiments P=0.0121), suggesting they may be predominantly excitatory; no differences in spike height or Fano factor (P>0.3).
• Single-unit robustness: After stringent single-unit classification, ESNs remained significant in Experiment 1 (95/373; P<0.001) and trended in Experiment 2 (21/132; P=0.0714).

The findings support the hypothesis that individual human hippocampal neurons can code the conjunction of elements that define a unique episode, consistent with the hippocampal indexing theory. ESNs reinstated their firing at retrieval for specific episodes, and this reinstatement could not be explained by selectivity to visual content (concept neurons) or by invariant timing (time cells). The presence of both rate-code ESNs and temporal-code tESNs, with significant episode-level overlap, suggests that hippocampal neurons may use complementary coding schemes to support episodic reinstatement. The lack of significant ESNs in parahippocampus (with limited sampling) aligns with a hippocampal-specific indexing mechanism. Although ESNs were not significant when analyses were restricted to later-forgotten events, bootstrapped comparisons indicated similar numbers of hit- and miss-ESNs, implying that neural reinstatement may occur without overt behavioral retrieval due to downstream limitations or encoding strategies. Wider spike waveforms among ESNs hint at excitatory cell involvement. Overall, the results provide single-neuron evidence that episodic memories are represented via sparse, distributed assemblies of hippocampal neurons that bind content and temporal context into unique episode codes.

This work demonstrates that human hippocampal neurons can act as episode-specific neurons that reinstate firing for unique episodes during successful memory retrieval. ESNs rely on rate and temporal firing codes, are largely confined to the hippocampus, and are distinct from concept neurons and time cells. The results support hippocampal indexing theory, whereby assemblies of ESNs serve as pointers that bind and later reinstate episodic memories. Future research should examine the stability of ESNs across repeated reactivations and longer timescales, determine regional and subfield specificity within the hippocampus, directly probe the network assemblies underlying episode codes, and assess sparsity and coding efficiency with paradigms allowing more precise control over timing and multiple reactivations.

• Each episode was encoded and retrieved only once, limiting reliability of single-neuron identification across repetitions and precluding assessment of ESN stability over time.
• The analysis framework, while controlled at the group level, does not allow firm claims about code sparsity (population or lifetime) due to per-episode testing without family-wise correction at the unit level.
• Parahippocampal sampling was limited (few microwire bundles and sessions), constraining conclusions about regional specificity.
• In Experiment 2, restricting to putative single units reduced statistical power (only a trend toward significance).
• The self-paced design and variable timing may have limited detection of time cells and precise temporal alignment of encoding/retrieval processes.
• Inability to assign units to hippocampal subfields prevents subfield-level interpretations.
• Potential for neural reinstatement without behavioral retrieval complicates interpretation of miss-ESN analyses.
• Visual tuning identification may miss weakly tuned concept neurons or inadvertently exclude reactivated ESNs during the tuning task, though analyses suggest minimal impact.