Hippocampal mismatch signals are based on episodic memories and not schematic knowledge

The brain uses prior knowledge to generate predictions and detect mismatches between expectations and current experience, which in turn drives learning. The hippocampus has been proposed to act as a comparator that detects novelty or mismatches by comparing incoming information against stored representations. However, it is unclear whether these comparisons are made against episodic memories of specific events or against generalized schematic knowledge abstracted across experiences. Prior work shows hippocampal sensitivity to changes in recently learned cue–outcome associations and event sequences (human and animal studies), consistent with episodic mismatch detection. Conversely, other accounts suggest the hippocampus also supports generalized predictive processes and statistical learning, raising the possibility that it could detect mismatches relative to schemas or abstract knowledge structures. The critical question addressed here is whether hippocampal mismatch signals are driven by violations of episodic memory, generalized schematic expectations, or both.

The paper reviews several lines of evidence: (1) Theoretical and empirical proposals that the hippocampus compares inputs with stored representations to detect associative and contextual novelty and schema incongruence, consistent with its specialization for relational and item–context processing and circuitry supporting comparator functions. (2) Strong evidence that hippocampus detects mismatches tied to specific prior experiences in humans (changes to cue–outcome relations and event sequences) and animals (environmental changes). Some computational models constrain hippocampal comparator operations to episodic-like representations. (3) Evidence for hippocampal roles in learning regularities, inferring unexperienced relationships, and imagining futures supports broader predictive coding/generative model accounts, leaving open whether schematic mismatches engage the hippocampus. The authors also situate findings within network-level perspectives: Default Mode Network (episodic integration over time), Semantic Control Network (context-appropriate semantic retrieval under uncertainty), and Multiple Demand Network (domain-general attention to surprising/difficult input).

Design: Three fMRI experiments manipulated the source of expectations while participants viewed custom stop-motion videos depicting everyday action sequences, with a single Target Action that was either Typical (contextually fitting) or Atypical (incongruent). By varying prescan familiarization, the authors isolated types of expectation violations at the Target Action: Experiment 1 (no prescan viewing): Atypical targets violated schema knowledge only. Experiment 2 (prewatch all Typical versions): Atypical violated both episodic memory for specific clips and schema knowledge. Experiment 3 (prewatch all Atypical versions): Typical violated episodic memory only (opposed to schema). Each participant viewed 34 clips in the scanner (17 Typical, 17 Atypical), each clip once. Participants: Exp1 N=36 (29F), 18–30 years (mean 21.6±3.3); Exp2 N=33 (22F), 18–32 (mean 22±3.4); Exp3 N=30 (25F), 18–29 (mean 21±2.8). Right-handed, normal/corrected vision, fluent English speakers. Ethics approved; paid £10/h. Stimuli: 34 scenarios, each with two matched versions differing only in the Target Action (Typical vs Atypical). Target Action segments averaged ~30 s. Procedure: Prescanning (Exp1: no familiarization; Exp2: view all Typical versions twice with recall in between; Exp3: view all Atypical versions twice with recall in between). Scanning: random presentation of 34 clips (17 Typical, 17 Atypical), with a trial structure fixation–video–fixation–comprehension question–fixation–odd/even judgment. Postscanning: cued free recall of actions in all scanned clips, focusing on Target Actions; audio recorded. MRI Acquisition: 3T Siemens Prisma, 32-channel head coil. Functional gradient-echo EPI, multiband factor 3 (TR=1520 ms, TE=28 ms, flip=75°, FOV=208×208 mm, 72 slices, 2 mm isotropic). Spin-echo fieldmaps with reversed phase encoding acquired each run. T1 MPRAGE (1 mm isotropic; Exp1-2: TR=2530 ms, TE=1.63 ms, FA=7°; Exp3: TR=2300 ms, TE=2.19 ms, FA=9°). Preprocessing described in SI. First-level GLM: Exp1 regressors: video onset, scene changes, Atypical target, Typical target, question+post-question fixation (target and other event regressors modeled as delta functions). Exp2–3 regressors: video onset, Atypical target, Typical target, question+fixations (no scene change regressor). Nuisance: 6 motion params, framewise displacement, white matter, CSF signals. ROI Analyses: Hippocampus and VTA/SN ROI: average beta weights for Atypical and Typical targets extracted (FSL fslmeants), two-tailed paired t tests; Bayes Factors for hippocampus (Cauchy prior scale 0.707). Network ROI analyses: Semantic Control Network (SCN), Multiple Demand Network (MDN), Default Mode Network (DMN); repeated-measures ANOVA with factors Condition (Typical vs Atypical) and Network, followed by Bonferroni-corrected paired t tests. Whole-brain: One-sample t tests on contrasts Atypical>Typical and Typical>Atypical; cluster-level FWE P<0.05 with voxel threshold P<0.001. Behavior: Logistic mixed-effects models (lme4) for correct recall (binary) with Condition and random intercepts for Participant and Clip; cross-experiment model included Experiment. Memory error analysis (non-preregistered): within forgotten/incorrect trials, coded 1 if action mentioned but object unspecified/incorrect vs 0 otherwise; model with Condition and random intercepts.

Behavioral memory: Across experiments, overall correct recall of Target Actions was high and did not differ between Typical and Atypical within each experiment (Exp1: β=0.12, 95% CI [-0.57,0.81], Z=0.33, P=0.739; Exp2: β=-0.59, 95% CI [-1.23,0.06], Z=-1.79, P=0.074; Exp3: β=0.42, 95% CI [-0.31,1.15], Z=1.14, P=0.254). Accuracy increased across experiments (β=1.23, 95% CI [1.09,1.37], Z=16.85, P<0.001), likely reflecting prescan familiarization. Participants made more memory errors for unexpected Atypical actions in Exp1 (β=1.51, 95% CI [0.77,2.25], Z=3.99, P<0.001) and Exp2 (β=2.48, 95% CI [1.61,3.34], Z=5.64, P<0.001); Exp3 showed a marginal reverse trend (B=-0.81, Z=-1.91, P=0.057), with more errors for Typical actions unexpected based on episodic memory. Hippocampus (primary ROI):

• Exp1 (schema-only mismatch): No hippocampal effect (paired t(35)=0.57, mean diff=0.14, 95% CI [-0.36,0.65], P=0.567; d=0.12, 95% CI [-0.30,0.54]). Bayes factor supported null (BF10=4.78).
• Exp2 (episodic+schema mismatch): Atypical>Typical, t(32)=-4.59, mean diff=-1.28, 95% CI [-1.84,-0.71], P<0.001; large effect d=-1.06, 95% CI [-1.64,-0.48]; BF10=378.25.
• Exp3 (episodic-only mismatch): Typical>Atypical, t(29)=5.00, mean diff=1.13, 95% CI [0.67,1.60], P<0.001; d=0.95, 95% CI [0.49,1.40]; BF10=910.34. A direct cross-experiment comparison (SI) showed a significant Experiment×Condition interaction indicating stronger hippocampal response to episodic than schema violations. Subregion and subfield analyses (head/body/tail; CA1, subiculum, CA3/CA4/DG) showed consistent effects throughout the hippocampus. Whole-brain and Network ROIs:
• Exp1 (schema-only): Atypical>Typical engaged attentional/semantic/predictive regions and subcortical caudate, amygdala, thalamus. Typical> Atypical engaged posterior medial and medial prefrontal regions (schema-consistent processing). Network ANOVA: Condition F(1,35)=13.70, P<0.001; Network F(2,70)=35.60, P<0.001; Interaction F(1.65,57.7)=14.07, P<0.001. Pairwise: SCN Atypical>Typical t(35)=5.64, P<0.001; MDN Atypical>Typical t(35)=4.11, P<0.001; DMN n.s. t(35)=0.30, P=0.763. VTA/SN: Atypical>Typical, t(35)=2.90, P=0.006 (Mdiff≈1.01).
• Exp2 (episodic+schema): Similar Atypical>Typical pattern including caudate, thalamus, amygdala. Network ANOVA: Condition F(1,32)=57.76, P<0.001; Network F(2,64)=24.87, P<0.001; Interaction F(1.29,41.26)=11.19, P<0.001. Pairwise: SCN t(33)=10.10, P<0.001; MDN t(33)=4.95, P<0.001; DMN t(33)=3.52, P=0.001 (Atypical>Typical). VTA/SN: Atypical>Typical, t(32)=5.97, P<0.001. Whole-brain interaction (Exp2>Exp1 mismatches) revealed right hippocampal selectivity for episodic-related violations.
• Exp3 (episodic-only): No Atypical>Typical clusters; Typical>Atypical engaged dorsomedial and inferior frontal cortex; subcortically, right amygdala and a small right ventral striatum region. Network ANOVA: Condition F(1,29)=19.50, P<0.001; Network F(1.33,38.59)=6.87, P=0.007; Interaction F(2,58)=6.56, P=0.003. SCN showed higher responses to unexpected actions generally, with significant Atypical>Typical t(29)=3.56, P=0.001; DMN showed Typical>Atypical t(29)=4.72, P<0.001; MDN not significantly modulated t(29)=1.29, P=0.207. VTA/SN: no significant difference t(29)=1.38, P=0.178. Summary: Hippocampal mismatch responses were absent for schema-only violations but robust for episodic-based violations (alone or combined with schema). Schema violations engaged SCN/MDN and subcortical PE-related regions (including VTA/SN), while episodic violations additionally recruited DMN. VTA/SN responses differed from hippocampus across experiments, emphasizing distinct roles.

The study directly tests whether hippocampal mismatch signals index violations of episodic memories or generalized schemas. Across three fMRI experiments, hippocampal responses were selectively elevated when ongoing events deviated from specific, recently encoded episodic memories (Experiments 2 and 3), but not when deviations were based only on generalized schema knowledge (Experiment 1). These results impose strong constraints on hippocampal comparator models, supporting accounts that limit mismatch detection to episodic-like representations and challenging broader theories proposing domain-general contextual mismatch detection (including schema incongruence and general predictive coding roles). The findings are not parsimoniously explained by global novelty, as hippocampal responses were similar in Experiments 2 and 3 despite differences in schema-based novelty, and absent in Experiment 1 despite schema-novel events. Network-level results reveal complementary roles: SCN and MDN support processing of surprising information and resolving ambiguity, particularly for schema-based violations; DMN engagement tracks episodic expectation violations, consistent with comparing internally generated (episodic) predictions with sensory input and potential reinstatement across the DMN. Subcortical VTA/SN showed prediction error-like responses to schema or combined schema+episodic violations but not episodic-only, diverging from hippocampal selectivity and suggesting different computations and interactions within hippocampus–midbrain loops. The authors propose that hippocampal mismatch detection applies to predictions situated within a learned cognitive map (spatial or conceptual), allowing flexible inference within that map, whereas mismatches derived from abstracted schematic models may be detected outside the hippocampus. Behaviorally, unexpected actions were remembered differently: while overall correct recall did not differ between expected and unexpected actions, unexpected actions elicited more memory errors (misremembered objects or unspecified objects), indicating that surprise alters the fidelity of memory content, potentially reflecting competition between memorability and contextual fit. These findings refine our understanding of how episodic memory supports online interpretation of unfolding events and how distinct neural systems contribute to learning from different types of expectation violations.

This work demonstrates that hippocampal mismatch signals are specialized for comparing ongoing experience to specific episodic memories, not generalized schematic knowledge. Schema-based violations recruit control networks (SCN, MDN) and subcortical prediction error circuitry, while episodic violations additionally engage the DMN alongside the hippocampus. The results support comparator models constrained to episodic representations and call for revisions to broader accounts positing domain-general hippocampal mismatch detection. Future research should determine whether highly precise generalized expectations (e.g., tightly structured routines like airport security) can elicit hippocampal mismatch signals without invoking a specific episodic memory, delineate the role of cognitive maps in guiding hippocampal comparisons across similar-but-novel contexts, and uncover cellular mechanisms (dopaminergic vs cholinergic modulation) underlying the hippocampal BOLD mismatch response.

The experiments focused on recently learned episodic expectations for specific video stimuli; generalization to other modalities, timescales, or more abstract schemas remains to be tested. Experiment 2 was not preregistered (though analyses matched other experiments). BOLD fMRI cannot disentangle underlying cellular mechanisms; midbrain (VTA/SN) and hippocampal divergences suggest complex neuromodulatory dynamics that were not directly measured. The design cannot fully exclude all novelty-related contributions, though analyses argue against a simple novelty account. Affordances of schema precision were not parametrically manipulated; thus, whether highly specific generalized expectations can drive hippocampal mismatches remains an open question. Finally, ROI/network definitions and whole-brain thresholds, while standard, may miss finer-grained or transient effects.