KurisutinaIdentity and status. Dagmar Zeithamova, April L. Dominick, and Alison R. Preston, Hippocampal and Ventral Medial Prefrontal Activation during Retrieval-Mediated Learning Supports Novel Inference. Neuron 75(1), 168–179. DOI; PMC author manuscript; published PDF linked by Preston's laboratory; publisher supplement. Published primary human fMRI experiment, online July 11 and issue dated July 12, 2012. Accessed 22 September 2026.
Factual core
Of 34 participants, 26 remained after imaging and learning-performance exclusions. Participants repeatedly studied overlapping image pairs and then chose directly associated or indirectly related images. A classifier trained on separate object/scene encoding data detected category-related activity during repeated learning when the associated image was absent. Controls matched visible category and exposure; a three-way classifier addressed the two-way output constraint. Reactivation change correlated with later inference accuracy across people (Spearman r=.46). Anatomical regional analyses related anterior medial temporal activity to reactivation and hippocampal/prefrontal changes to inference. Controlling premise accuracy retained the ventromedial prefrontal association; the bilateral hippocampal association was nonsignificant, with a weaker right-only result. Hippocampal–prefrontal coupling increased across repetitions. The study supports learning-phase reactivation associated with inference; item-specific recovery, causal integration, and permanent storage were not directly measured. Main pp. 168–177 and supplement S1–S11.
Decoder and inference audit
The prediction target has three different levels. Held-out localizer runs test discrimination of visible object/scene categories. Guided recall tests discrimination of instructed recall category in the same people, using different runs. Applying that classifier to the associative task measures a category-sensitive signal when related content is absent. These are useful separations of training and application, but none is a held-out forecast of a new person's inference performance. The reported behavioral relationship is a correlation estimated across the analyzed people. Neither a category score nor the illustrative recall timecourses identify which particular image was retrieved, its exact features, or whether the person preserved its source.
The supplement supplies meaningful validation: four-fold leave-one-localizer-run-out classification averaged 94%; encoding-trained classification during guided recall averaged 69%. Localizer/recall stimuli differed from the associative-task stimuli, and that session preceded the main experiment by 1–7 days. Guided recall used category instructions and a key press for each claimed recall, rather than an independently checked item report. Thus category-directed imagery or semantic activity remains compatible with the validation endpoint; describing it as recovery of verified event details would go beyond the measurement. The logistic output is not independently calibrated as a probability that a particular memory is accurate. Supplement S7–S9.
Useful controls do not establish the storage mechanism. The same visible-category comparison and matched repetition counts constrain a simple novelty explanation. The three-way analysis addresses the mechanical increase in one category's score when the other decreases in a two-way classifier. Anatomical masks avoid choosing the main classifier voxels by their correlation with inference, and excluding anterior MTL voxels from the reactivation measure retains its regional correlation. These controls strengthen a category-reactivation interpretation; they do not decode a durable A–C binding. Similar aggregate accuracy across conditions is also not proof that their attention, strategy, and difficulty are identical.
Partial adjustment is not elimination of the premise-memory account. The exact estimates matter: VMPFC partial r=.53, p=.007; bilateral hippocampus partial r=.22, p=.29; right hippocampus partial r=.39, p=.05. The right-only result should not be silently substituted for the bilateral result or described as a corrected whole-hippocampus effect. Corrected significance is explicitly reported for selected analyses, not for every exploratory region, hemisphere, and comparison. Moreover, conditioning on a short, high-performing recognition score need not condition on all relevant associative strength. Main p. 172.
Original mathematical illustration, not the paper's fitted model. Let latent premise strength L have variance 1, and stipulate X=L+u, Y=L+v, and observed premise score P=L+e, with mutually independent zero-mean Gaussian L,u,v,e and Var(e)=s²>0. Then
Cov(X,Y | P) = 1 − 1/(1+s²) = s²/(1+s²) > 0.
Here a positive adjusted association between neural measurement X and inference Y arises without an integration variable. This establishes a possible alternative, not the actual explanation of the results. The paper uses rank-based analyses; this deliberately simple linear-Gaussian example is not a numerical reanalysis of its correlations.
Dependence and selection require separate judgments. Excluding three participants below 75% premise accuracy limits the analyzed population to relatively successful learners; four motion exclusions and one scanner-artifact exclusion are also reported. The localizer's run-based split is stronger than randomly splitting adjacent scans. However, the supplement's nominal individual above-chance count uses an independent-binomial calculation on timepoints within blocked fMRI runs. Adjacent scans are not independent Bernoulli replications, so that count needs a block-aware null before being treated as an individually validated detector. This concern is distinct from the participant-level group analyses and does not by itself invalidate them.
Two unresolved reporting discrepancies. Main Methods says 26 participants enter every reported analysis, but connectivity tests repeatedly use F(1,21), including Figure S4. Neither retained Methods section explains this denominator; do not invent four additional exclusions. The test-count sentence also says 16 AB, 16 BC, and 16 AC trials for each triad type after a run, while the stated schedule implies 16 total triads per run, four of each type. Forty-eight total tests per run is the coherent reading, but the original task script was not inspected. These are reporting limits, not evidence of misconduct or proof that the main effects are incorrect.
Original interpretation in relation to REMERGE
Learning-phase reactivation is informative against a strict account in which related content is never retrieved before the final query. It does not refute all retrieval-time inference accounts. A recurrent learner could retrieve related episodes during encoding, strengthen or reorganize their separate representations, and still compose the final answer at test. Alternatively, it could store a derived association. Both can produce reactivation during learning and good later inference. The REMERGE reading supplies a constructive retrieval alternative but does not itself simulate this experiment's trial-by-trial learning and measurement process.
Functional coupling also does not reveal transfer direction, a unique binding operation, or the destination of permanently stored content. Task and motion regressors help address shared variance; interpreting the remaining coupling as one region writing a particular memory into another requires additional evidence. The absence of an overall run trend does not establish that repetition effects occur only for overlapping events, because a matched nonoverlapping associative-learning condition is missing from that contrast.
The inference task was explained in advance, practiced, and repeated after each run. This supports a claim about inference under those learning goals; spontaneous autobiographical integration under unknown future demands is a separate target. Testing each A–C relation before its premise probes protects against those particular probes having already constructed the inference, but the inference test itself can still retrieve, compute, and subsequently store an answer. Correct final choice therefore does not timestamp when the answer first became stored.
Discriminating follow-up and transfer implications — proposed, not performed
First sharpen the neural target. Use independent localizers and held-out runs to distinguish the associated C item from other same-category items. Add evaluator-held arbitrary details and source attributes that category knowledge cannot supply. Compare decoding against category-only and task-instruction baselines; use block-preserving nulls. Item-specific evidence would strengthen content reactivation, but still would not prove that a new integrated trace was stored.
To test storage, compare explicit implementations that retrieve separate episodes, cache inferred relations, or do both. Give them identical base records, then change a premise after learning. Prespecify which inferences should update, which original details should survive, how response time or compute should change, and whether each architecture can recompute a cached answer. Verify learning of the changed premise and include matched nonconflicting additions. Selective retrieval restrictions can distinguish specified artificial implementations; they cannot serve as a universal diagnostic of biological episodic versus integrated memory.
For a person model, separately evaluate future item-level responses on held-out episodes, calibrated predictions of that person's errors, and the recipient's own decisions requiring a new binding. Preserve whether an assertion was directly observed, inferred, or later remembered as observed. Successful inference with weak source discrimination can be useful reasoning while still being an incomplete transfer of personal memory. A neural category decoder should not be credited with supplying arbitrary episode details that entered through the stimulus list, prompt, or evaluator key instead.
Reading and archive record
Read the full PMC author-manuscript main text, then all 12 pages of the published main PDF, including Methods, Discussion, six figures/captions and references. Read the complete 11-page publisher supplement: four figures/captions, all supplemental procedures, and two references. No tables occur. Visually inspected main PDF pages 3–9 and supplemental pages 2, 3, 5, 6, and 9, covering every figure, classifier controls, partial correlations, participant selection, and the reporting discrepancies. No graph digitization, original code execution, raw-data analysis, or later replication review was performed.
Retained versions are identified separately: author-linked final publisher PDF; PMC accepted-manuscript HTML; publisher supplemental PDF. Complete byte/text equivalence is not presumed. Direct PMC PDF/DOCX requests returned HTML challenges, and the Europe PMC supplement API reported that this manuscript was outside its open-access download set. Valid public author/publisher copies were obtained instead with normal TLS. The PMC DOCX itself was not acquired or compared with the publisher supplement. Copyright remains with Elsevier; no open reuse license was identified.
Local published PDF, PDF text, PMC HTML, PMC reading text, supplement PDF, supplement text, and provenance/hashes.