KurisutinaResearch date: 2026-09-22.
Citation and status. Alexa Tompary and Lila Davachi. Consolidation Promotes the Emergence of Representational Overlap in the Hippocampus and Medial Prefrontal Cortex. Peer-reviewed, Neuron 96, 228–241.e5; published 2017-09-27. Article DOI; public author-hosted publisher PDF.
Acquisition and reading. Saved PDF, readable text, and provenance with SHA-256 hashes. Read the complete main article, all figure legends, references, and STAR Methods e1–e5. Separate Supplemental Information, containing four figures and seven tables, was not obtained or read; main-text descriptions of its controls were read. The saved version is the publisher article hosted by an author's laboratory; no open reuse license was identified.
Source findings and limits (factual paraphrase). Twenty-two adults learned 128 object–scene pairs sharing four scenes; nineteen entered similarity analyses. Different pairs were tested immediately or after one week. Scene-recall accuracy fell from 94.2% to 54.1%. Among confidently and correctly recalled pairs, remote retrieval patterns reflected shared scenes in medial prefrontal cortex and hippocampal subregions; the remote comparison pooling the whole hippocampus was not significant. Remote hippocampal reinstatement of individual episodes coexisted with overlap, and stronger reinstatement related to weaker overlap representation. This was not a behavioral test of novel generalization. The authors discuss selective forgetting, changing retrieval strategies, and selection of stronger surviving remote memories as alternatives to integration. Post-encoding connectivity correlated with remote overlap, but its relationship was not significantly stronger than for recent overlap. Thus the results are consistent with memory transformation, without isolating consolidation from elapsed time or proving that abstract knowledge improved, episode content was erased, or hippocampal–cortical interaction caused either change.
Interpretation for functional transfer — our reasoning. Three targets should remain separate: fidelity to a particular episode, access to that episode now, and useful generalization across episodes. Preserving the first does not guarantee the other two. Nor does a more compressed or generalized model necessarily resemble its source person more closely. A system can become better at inference while becoming worse at that person's exceptions and source distinctions.
The study motivates a dual representation as an engineering hypothesis: immutable episode evidence plus revisable regularities and a state-dependent retrieval process. It does not imply that a successful implementation must reproduce these brain regions, or that detail and generalization inevitably compete. Representational overlap can arise because shared information strengthened, distinctive information weakened, or the tested subset changed. The same observed similarity statistic does not distinguish those mechanisms.
Falsifiable experiment — proposal, not an existing result. Teach controlled episodes containing repeated regularities, source-specific details, and deliberate exceptions. Randomize delay and initial learning strength, assigning separate matched item families to each test occasion. Score all assigned trials; do not define the delayed analysis population solely by successful recall. At each occasion, separately measure episode/source discrimination, responses to unfamiliar cases requiring generalization, and false generalization over exceptions. Test whether a person-specific transition model forecasts the joint outcome distribution better than episode retrieval alone, learned rules alone, and a static combination supplied with the same evidence. Compare against simple strength and recency baselines. A model that predicts gist accuracy but misses exception errors has captured only one aspect of the person.
Independent item sets reduce direct retest effects but cannot guarantee independence when sets share rules or contexts. Counterbalance the relevant families or include a group tested only after the delay. Delay alone does not identify consolidation; stronger mechanistic claims require an additional intervention and controls for its other effects.
Connection to extraction. A late account cannot be treated as a transparent readout of the original episode. A transfer dataset should preserve dated observations of recollection and distinguish evidence about the event from evidence about how it is currently reconstructed. Functional evaluation should ask what the system predicts correctly as this balance changes.