KurisutinaAidan J. Horner, James A. Bisby, Daniel Bush, Wen-Jing Lin, and Neil Burgess. Evidence for holistic episodic recollection via hippocampal pattern completion. Nature Communications 6, 7462. Published 2 July 2015. Peer-reviewed human experiment with a computational model. Publisher, PMC, DOI: 10.1038/ncomms8462.
Evidence from this source
Twenty-six healthy adults learned 36 synthetic events through interleaved pairs. Closed loops connected three elements pairwise; open chains linked four elements through three pairs. Tests used six-choice associative recognition. Closed loops produced greater dependence among retrieval successes and stronger incidental category-specific cortical activation; this activation covaried across participants with hippocampal activity. A recurrent associative model reproduced the predicted pattern. The findings support binding and completion beyond independent pair retrieval. They do not decode specific incidental items, observe synaptic engrams, establish hippocampus-to-cortex causation, or extract unknown autobiographical memories. Neural reinstatement was measured at the level of categories. Three-versus-four elements remain a design difference, although the authors discuss prior controls. Overall performance was used to set the model's response threshold; this was not a parameter-free prediction. Authors frame the all-to-all structure as a possible boundary condition, not a universal requirement for episodic retrieval.
Original interpretation for memory extraction and functional transfer
The transferable target should include relations among retrievals, not only a database of retrievable facts. Two systems could remember the same number of facts yet differ in whether recalling one detail automatically brings related details to mind. That difference matters for an Amadeus-like system: spontaneous associations may shape the next interpretation or decision even when the user did not request them.
An association graph is a possible engineering representation of such relations. This paper does not select that architecture over a learned latent representation or a retrieval process that operates over flat records. Its computational account is a demonstration of sufficiency under chosen assumptions, not identification of the one algorithm used by the brain. A new system need not reproduce biological recurrence to reproduce the relevant behavior; conversely, implementing recurrence does not establish a functional memory copy.
The word “holistic” should not imply exhaustive or accurate playback. For this program, the useful claim is narrower: partial cues may recruit multiple related components, and the pattern of recruitment can be measured. An interviewer who acquires only explicitly requested details may miss those relations. Acquiring a coherent narrative can also impose relations that were not present, so extraction and construction need separate controls.
Proposed discriminating experiment
Hold a known event inventory and marginal retrieval accuracy approximately fixed. Compare a model given facts alone with models given observed cue-to-detail links or joint retrieval data. Use new cues and score joint retrieval distributions, incidental intrusions, and downstream choices—not merely whether each fact can eventually be prompted. Include a simple associative-network comparator and an equal-information full-context comparator. Withhold some relations entirely for transfer testing. Counterbalance probe order and estimate retrieval-induced carryover rather than assuming every probe measures an unchanged trace.
Success would establish a useful functional representation on those events. It would not show that the representation is the person's hippocampal engram or that it generalizes to remote, naturally formed memories.
Reading and acquisition record
Read the complete main introduction, results, discussion, methods, model description, and figure captions from the public full-text XML; checked the published PDF and its textual model equations. Supplementary Figure 1, Tables 1–4 and Notes 1–5 were not read. Graphical figures and raw data/code were not independently audited; no reanalysis was performed.
Local sources: published PDF, Europe PMC XML, readable text, provenance and SHA-256 hashes. Text extraction retains prose and captions; some equations are stored as graphics in the XML, so consult the PDF for those expressions. The paper is CC BY 4.0, subject to indicated third-party exceptions.