KurisutinaAdrian W. Gilmore, Alina Quach, Sarah E. Kalinowski, Estefanía I. González-Araya, Stephen J. Gotts, Daniel L. Schacter, and Alex Martin. Evidence supporting a time-limited hippocampal role in retrieving autobiographical memories. PNAS 118(12), e2023069118. Published online 15 March 2021. Peer-reviewed primary human fMRI experiment. DOI: 10.1073/pnas.2023069118. Public PMC full text; author-institution PDF.
Evidence from this source
Forty healthy young adults narrated six memories per period: today, 6–18 months, and 5–10 years earlier, approximately two minutes each. Transcript details were classified and time-locked to fMRI; regressors separated detail-associated transients from sustained recall effects. Posterior, not anterior, hippocampus showed stronger sustained activity for today than remote memories. Remote posterior activity did not significantly exceed a picture-description control. Authors interpret this as supporting standard consolidation. Remote narratives contained somewhat fewer internal details, but no significant change in the internal/external ratio; content categories changed differently. External details included other-event episodic material, repetitions, and semantic content. Independent readers (129 vividness raters; 144 detail raters) showed no significant age-related rating differences. Right posterior hippocampal connectivity with scene-selective cortex also declined with memory age. This was retrospective sampling, not longitudinal observation or lesion evidence of necessity. The control described a just-viewed picture from memory. Different cue subsets were reserved for today's events. Main paper and linked supplement.
Original interpretation and competing explanations
This is useful evidence against assuming that all rich remote recall has an identical aggregate hippocampal activation profile. It does not establish that the hippocampus is dispensable. Failure to reject a remote-versus-baseline difference is not an equivalence test, and equal aggregate activity can coexist with different information processing. The imaging result supports the authors' operational predictions without uniquely choosing a theory of storage or necessity.
The distinction between transient detail effects and sustained recall effects changes the question being asked. It estimates temporal differences conditional on a model of narrated content, rather than total hippocampal participation in remembering. If a retrieval mechanism works through constructing details, accounting for that component can remove part of the mechanism's contribution. This possibility does not invalidate the observed residual age gradient; it prevents reading that gradient as a direct assay of the entire memory trace. Baseline selection likewise matters: a control useful for matching speech and encoding can share some of the operations whose necessity is disputed.
Reader judgments about transcripts cannot establish equivalence of the original rememberers' phenomenal experience. Internal-detail counts, internal/external ratios, and subjective vividness should therefore remain separate endpoints. An external-detail count is not a pure semantic-memory measure, so an unchanged ratio cannot by itself exclude transformation toward gist. Narrative detail can reflect communication style, event selection, or reconstruction as well as retained information. Retrospective age bins also conflate elapsed time with the kinds of events that survive selection, original developmental period, and rehearsal history. These competing explanations require measurement or manipulation, rather than being settled by changing the theory label.
Proposed experiment for functional transfer
Preserve both event-specific records and abstracted knowledge, then test their contributions separately. Cross old/recent events with independently assessed specificity and rehearsal, and compare performance under open narration, targeted factual cues, and novel cue combinations. Report verified detail accuracy, source binding, cue sensitivity, and downstream decisions alongside richness and similarity of wording. Avoid interpreting a system's ability to generate an old story as evidence that the underlying retrieval organization transferred.
For mechanistic models, predefine both the total effect of memory age and the effect remaining after content adjustment. Test whether conclusions change across baselines with different demands for speech, scene construction, and new encoding. For functional extraction, an important criterion is whether neural data improves held-out prediction beyond matched behavioral information. Mechanistic evidence can have value without that incremental forecasting result. A hippocampal activation map alone does not specify a transferable autobiographical representation.
Comparison with Steinvorth et al. (2005)
The lesion study concerns the availability of supported autobiographical detail after broad tissue damage; this imaging study concerns modeled activity during successful recall in intact young adults. Their populations, sampled memory ages, retrieval support, and endpoints differ. They are not a controlled head-to-head adjudication. Together they motivate a distinction between event knowledge, access to specific details, and the operations used during recall. Neither licenses an age threshold after which an event can safely be replaced by gist in a functional memory model.
Reading and acquisition record
Read the complete main introduction, results, discussion, methods, and figure captions from the public XML and checked the publisher-formatted 12-page PDF. Also read all supplementary methods, Tables S1–S3, and captions of Figures S1–S5 from the 16-page supplementary PDF. This includes exclusions, scoring/reliability, reader-rating instructions, imaging preprocessing, and the residual-detail control analysis. Supplement references were scanned, not followed individually. Graphical values were not independently remeasured; raw data, code, and individual behavioral narratives were not inspected or reanalyzed.
Local sources: published PDF, Europe PMC XML, readable text, supplement PDF, supplement text, provenance and SHA-256 hashes. The article's stated license is CC BY-NC-ND 4.0. Accessed 22 September 2026.