{"id":"0f203508-07f0-4ae4-9c01-f47b9676dcb6","arxiv_id":"2501.02950","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The brain may separate memory addresses (keys in the hippocampus) from memory content (values in the neocortex), making forgetting a retrieval failure rather than a storage failure.","lead":"This paper proposes that the brain uses key-value memory: the hippocampus stores keys (addresses) and the neocortex stores values (content). The authors review psychological and neural evidence and run two toy simulations that illustrate retrieval failure and recovery.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The key/value division is underdetermined: the paper offers no operational criterion for classifying a hippocampal representation as a key rather than content, so the cited evidence cannot establish the central anatomical claim.","rationale":"The reader's conditional verdict is appropriate. The most load-bearing assumption is the anatomical key/value mapping, because the paper's novelty and title depend on it. I agree with the reader that the evidence is correlational, and I would sharpen the concern: there is no operational criterion for classifying a representation as a key rather than a value. The paper's own formal framework is broad enough that any content-addressable code can be described as a key, which makes the division of labor unfalsifiable as presented. The paper acknowledges its speculative status, which is to its credit, but that does not remove the need for a discriminating test. I therefore recommend leaving the verdict unchanged: the framework is plausible and worth developing, but the central anatomical claim requires the repulsion-reversibility test (or a similarly direct dissociation) before it can be regarded as empirically supported. I do not see an internal inconsistency in the simulations; the beta-dependence noted by the reader is a secondary concern, not the load-bearing issue.","tokens_in":19762,"tokens_out":9386,"duration_ms":92619,"concrete_test":"Run the paper's own untested prediction from Section 6: in the Chanales et al. (2017) overlapping-routes paradigm, measure hippocampal representational repulsion after discrimination training, then remove the discrimination requirement and measure repulsion again. If repulsion does not reverse, the claim that hippocampal keys are optimized for discriminability is falsified, and the central division-of-labor claim loses its key positive support. If repulsion does reverse, the key-value interpretation is supported over fixed content stores.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that the hippocampus stores keys while the neocortex stores values—requires hippocampal representations to function as addresses rather than content. But the evidence marshaled in §4.2 (sparse conjunctive engrams, reinstatement, optogenetic reactivation, the Chettih et al. food-caching barcodes) is equally consistent with the hippocampus storing the episodic content itself. The paper itself notes in §3 that Hopfield networks and the Tolman-Eichenbaum Machine use the same representation as both key and value, and in Box 1 the MESH/Vector-HaSH models place keys in the hippocampus while values are reconstructed at the sensory layer; neither source provides a dissociation. No formal or empirical criterion is offered to determine whether a neural population is a key store or a value store: any content-addressable code can be labeled as a key within this framework. Thus the proposed division of labor is a reinterpretation of known phenomena rather than a tested hypothesis. The only novel falsifiable prediction offered (Section 6: repulsion effects in long-term memory should be reversible) has not been tested, and it targets the 'keys optimized for discriminability' entailment, not the anatomical mapping directly. If hippocampal representations carry content, or if repulsion is not reversible, the specific claim fails. The paper labels itself speculative, which is honest, but the central anatomical claim rests on this underdetermination.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper argues that memory in the brain can be understood as a key-value memory system, with the hippocampus serving as a key/address store and the neocortex as a value/content store. It formalizes key-value memory through correlation matrix memories and the dual form of linear layers, surveys psychological and neural evidence for retrieval-oriented forgetting and for distinct keys and values, and presents two simulations: one showing that key and value representations diverge under optimization, and another showing that amplifying stored keys for a 'forgotten' task can restore performance after continual learning. The authors explicitly label the central anatomical proposal speculative and offer one untested prediction about reversibility of repulsion effects.","tokens_in":20058,"tokens_out":4611,"duration_ms":46902,"significance":"The mathematical core of the paper is correct and clearly presented: the dual-form equivalence for linear layers (Section 2.3, Eqs. 10-11) and the kernel formulation (Section 2.1) usefully connect Hopfield networks, sparse distributed memory, linear attention, and transformer self-attention. The review of behavioral evidence for retrieval-based forgetting is informative, and the public code release is a strength. However, the central claim of a hippocampal-key/neocortical-value division of labor is underdetermined by the evidence cited in Section 4.2, and the supporting simulations are too limited to serve as more than illustrations. The paper is best read as a perspective piece; as a scientific claim it needs a sharper operational criterion and a direct test.","major_comments":[{"comment":"The central anatomical claim—that hippocampal representations are keys and neocortical representations are values—lacks an operational criterion for classifying a neural representation as a key rather than as content. The evidence listed in §4.2 (sparse conjunctive engrams, hippocampal-dependent reinstatement, optogenetic reactivation, Chettih et al.'s food-caching barcodes) is equally compatible with the hippocampus storing the episodic content itself or an index that is itself a form of content; the paper itself notes in §3 that Hopfield networks and the Tolman-Eichenbaum Machine use the same representation as both key and value, and Box 1's Vector-HaSH places keys in the hippocampus while values are reconstructed at the sensory layer. Without a definition that would be violated if hippocampal codes carried content, the proposed division of labor is a reinterpretation of known phenomena rather than a tested hypothesis, and this bears directly on the paper's central claim.","section":"§4.2"},{"comment":"The continual-learning simulation is reported as a single trajectory with no error bars or multiple seeds, and the recovery effect is produced by scanning a single scalar β ('optogenetic strength') and selecting a value that improves Task 1 accuracy; this is a post-hoc demonstration rather than a predictive test. The paper also does not specify how β would map onto a real neural intervention or how the result distinguishes retrieval recovery from a generic increase in key-norm magnitude. With one run and a tuned parameter, the simulation does not substantiate the stronger claim that the model 'resonates with' silent engram recovery; it is illustrative only.","section":"§5.2, Fig. 3B"},{"comment":"The only explicitly novel falsifiable prediction, that repulsion effects in long-term memory should be reversible, has not been tested and targets the 'keys optimized for discriminability' entailment rather than the anatomical key/value mapping. The authors should either derive a test that would discriminate the hippocampal-keys/neocortical-values division from a content-storing hippocampus, or state plainly that this mapping is a non-exclusive proposal; as written, the central claim is not yet disconfirmable.","section":"§6"}],"minor_comments":[{"comment":"There are several typos and spelling inconsistencies, including 'inacessible' in §4.1, 'demantia' in §4.2, the section header 'V alues' in §4.3, 'Consisent' and 'repsectively' in §3 and Box 1, 'Tzyulmankov' in §3, and 'scaler' for 'scalar' in §5.2.","section":"Throughout"},{"comment":"Figure 2 shows trajectories for a single run without error bars or multiple initializations; the caption should state whether the displayed configuration is representative across seeds and initializations.","section":"§5.1, Fig. 2"},{"comment":"The abstract and introduction state that the paper presents 'simulations that recapitulate a number of empirical phenomena,' but only two toy simulations are presented; this wording is stronger than what the simulations support and should be tempered.","section":"Abstract/Introduction"},{"comment":"The box defines the key and value matrices as K = W_top^ad and V = W_ds but does not explain how a reader should reconcile this with the earlier definition of keys as stored address vectors and values as content; a small schematic or verbal explanation would improve accessibility.","section":"Box 1"},{"comment":"The metamemory evidence (feeling of knowing, tip of the tongue, change detection) is consistent with a separate key store but also with partial retrieval of content; adding a discussion of conditions that would distinguish these alternatives would strengthen the section.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a perspective/review that leans heavily on the authors' own prior work (Irie et al., MESH/Vector-HaSH) for the key architectural claims, which is acceptable for a perspective but the central anatomical claim needs independent operational support. The paper is honest about its speculative status, which mitigates concerns, but the underdetermination of the key-vs-content distinction is the main obstacle to publication as a claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful synthesis, not a breakthrough test. The genuinely new piece is applying the explicit key-value formalism to the hippocampus-neocortex division, and the one testable prediction—that repulsion effects in long-term memory should be reversible—is worth taking seriously. The rest of the paper is a well-organized review of ideas that already existed (hippocampal indexing, CLS, transformer attention).\n\nIt does several things well. The formal foundations are correct: the dual form of linear layers is a theorem, and the kernel framing is standard. The authors are upfront that the connections are speculative, and they release code for the simulations, which is more than many theory papers do. The citation pattern is fair; there is some self-citation (Irie et al., MESH/Vector-HaSH), but the framework does not depend on those particular results being true.\n\nThe soft spots are real but not disqualifying. The two simulations are toys. The first shows that gradient descent separates keys and values in a 2D toy problem; that is known behavior. The second shows that scaling keys by beta can restore performance on a forgotten task in a small MLP, but it appears to be a single run, no error bars, and beta is a post-hoc knob. More importantly, the central anatomical claim—hippocampus stores keys, neocortex stores values—is underdetermined. The cited evidence (sparse engrams, reinstatement, optogenetic reactivation, food-caching barcodes) is equally consistent with the hippocampus storing episodic content itself. The paper mentions that the Tolman-Eichenbaum Machine uses the same representation for key and value, but doesn't give an operational way to dissociate the two in neural data. So the division of labor is a reinterpretation, not a demonstrated fact. That's fine for a hypothesis paper, but a reader should not come away thinking the mapping is established. I'd like to see an empirical criterion—something like a prediction that a hippocampal manipulation can disrupt retrieval without degrading the content, or that key-space geometry changes under retrieval demands independently of value-space geometry.\n\nBottom line: the paper is for cognitive neuroscientists and ML researchers who want a compact bridge between transformer attention and hippocampal memory. It deserves serious peer review; I would send it in, with the expectation that the simulations get strengthened and the anatomical claim is framed more carefully as a proposal with a clear falsification.","headline":"A well-written synthesis that reframes hippocampal indexing as key-value memory; the central anatomical claim is underdetermined, but the paper is honest and deserves serious peer review.","tokens_in":20527,"tokens_out":3843,"would_cite":true,"duration_ms":88739,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that human memory is a key-value store: the hippocampus keeps the addresses (keys), the neocortex keeps the content (values), and forgetting is a retrieval failure, not an erasure.","keywords":["key-value memory","hippocampal indexing","memory retrieval","forgetting","associative memory","transformer attention","engrams","complementary learning systems"],"falsifier":"If a study found that a specific memory could be behaviorally recalled in full detail while its hippocampal engram was optogenetically silenced—or, conversely, that stimulating a hippocampal engram alone reproduced the memory's content without any neocortical reactivation—the proposed division of labor between hippocampal keys and neocortical values would be contradicted.","tokens_in":1771,"feed_emoji":"🧠","tokens_out":1779,"duration_ms":100078,"temperature":0.7,"pith_summary":"This paper argues that memory in the brain works the way a key-value store works in computing: experiences are split into two representations, keys that serve as addresses and values that hold content, and recall happens by matching a query to keys and then reading out the associated values. The proposed biological division of labor places keys in the medial temporal lobe (notably the hippocampus) and values in the neocortex. On this view, memories are rarely destroyed; they are stored indelibly and become unavailable because retrieval fails, which is why amnesia can shrink spontaneously and supposedly lost memories can be recovered with the right cues. The paper also argues that keys are optimized for discriminability, values for fidelity, and keys are not open to conscious recall. If this is right, it unifies disparate findings—from hippocampal engrams and cortical reinstatement to tip-of-the-tongue states and machine-learning attention—under one computational principle.","feed_headline":"Store memories with keys: hippocampus addresses, cortex content","feed_subtitle":"The hippocampus stores addresses, the neocortex stores content; forgetting is a retrieval failure, not erasure.","key_machinery":"The central object is the key-value memory, formalized as an association matrix updated by Hebbian outer products, $\\Delta M \\propto k_n^\\top v_n$, and read out as $\\hat v = qM$. Rewriting the readout in dual form, $\\hat v \\propto \\sum_n \\alpha_n v_n$ with attention weights $\\alpha = \\sigma(S(K,q))$, shows that any such memory is a soft retrieval over stored values weighted by query-key similarity; choosing different similarity kernels $S(\\cdot,\\cdot)$ and separation operators $\\sigma(\\cdot)$ yields Hopfield networks, sparse distributed memory, dense associative memory, linear attention, and transformer self-attention. The identity that carries the argument is that a linear layer trained by gradient descent is exactly a key-value memory, $y = xW_0 + \\sum_n \\alpha_n v_n$, with keys equal to layer inputs and values equal to error signals. The biology is carried by proposed implementations: Hebbian-like learning rules in the hippocampus for keys, cortical plasticity for values, and fixed random or structured scaffolds (for example, grid-cell-like attractors) that prespecify well-separated addresses.","core_discovery":"On the authors' account, the brain implements a key-value memory system. Each experience is encoded twice: a key vector that functions as a retrievable address, and a value vector that stores the content to be recovered. Retrieval is the soft match between a query and the stored keys, with the final readout a weighted combination of values; this is the same dual-form computation that underlies transformer self-attention, and the authors show that even an ordinary gradient-trained linear layer can be rewritten as a key-value memory whose values are the training errors it experienced. They posit that the hippocampus stores the keys and performs query-key matching, while the neocortex stores the values, with hippocampal 'engram' cells acting as causal indices that reinstate cortical content. This architecture predicts that memory traces are not erased by new learning, that hippocampal keys are shaped to discriminate overlapping experiences (repulsion), and that the keys themselves cannot be consciously recalled, even though they can support feelings of knowing and tip-of-the-tongue judgments. Two toy simulations illustrate the distinctive predictions: keys and values evolve toward different optimal geometries, and a network that has 'forgotten' a first task can recover it by amplifying the first task's stored keys, mirroring optogenetic reactivation of silent engrams.","pith_inferences":["Editorial inference: the reversible-forgetting account suggests a clinical direction—memory rehabilitation could target retrieval conditions rather than restorage, for example by designing cues that re-engage the original hippocampal address pattern.","Editorial inference: the key-value view predicts that individual differences in metamemory accuracy should track the quality of key-query matching, which could be measured with neural pattern-similarity analyses of hippocampal activity during failed recall.","Editorial inference: the fixed-scaffold models imply that learned keys are not always better; choosing the scaffold geometry could be as important as plasticity, a principle that may transfer to continual-learning systems.","Editorial inference: if keys are unavailable to conscious recall, representational analyses should find information in hippocampal activity that predicts successful retrieval but is not decodable in explicit report—an untested prediction."],"forward_implications":["Forgetting is retrieval failure: information is retained but inaccessible, so memory can be restored by improved cues, repeated retrieval attempts, reminder exposures, or direct reactivation of silent engrams.","Hippocampal representations should be shaped by discriminability demands, producing repulsion of overlapping experiences, whereas neocortical representations should preserve content fidelity; the paper predicts repulsion effects should be reversible.","Metamemory judgments such as feeling of knowing, tip-of-the-tongue states, and change detection without identification can be driven by key-query match alone, without value retrieval.","Machine-learning architectures that separate keys from values, including fixed scaffolds, can avoid catastrophic forgetting and show graceful degradation, unlike autoassociative models that mix keys and values.","If the hippocampus stores keys, then hippocampal damage should cause overgeneralization and loss of targeted access, with cortical values still intact."],"supporting_citations":[{"why":"Supplies the original correlation matrix memory model, the earliest formal key-value memory with a Hebbian outer-product update.","marker":"14"},{"why":"Introduces transformer self-attention, which the paper identifies as a key-value memory with softmax separation.","marker":"7"},{"why":"Provides the theorem that gradient-trained linear layers are equivalent to key-value memories with keys as inputs and values as error signals.","marker":"12"},{"why":"Proposes hippocampal memory indexing theory, the direct precursor of the claim that the hippocampus stores keys.","marker":"103"},{"why":"Proposes biologically plausible learning rules for keys and values in a three-layer key-value network, including least-recently-used key updates.","marker":"49"},{"why":"Shows that a fixed grid-like scaffold stores keys in hippocampus and reconstructs values at the sensory or cortical layer, outperforming a learned encoder.","marker":"55"},{"why":"Demonstrates that engram cells persist and can be optogenetically reactivated after retrograde amnesia, supporting retrieval failure over erasure.","marker":"129"},{"why":"Shows that a reminder restores context-specificity in normal rats but not hippocampal-lesioned rats, evidence that the hippocampus stores keys.","marker":"101"},{"why":"Reports repulsion of overlapping hippocampal representations during learning, evidence that hippocampal keys are optimized for discriminability.","marker":"112"},{"why":"Dissociates memory accessibility from precision over retention, showing forgetting reflects loss of access rather than degradation of stored content.","marker":"74"}],"fun_headline_variants":["Brain uses key-value memory: hippocampus addresses, cortex content","Forgetting is failed retrieval, not erased memory","Hippocampus stores keys, neocortex stores content","Key-value memory explains why forgetting is a retrieval failure","Brain's key-value store: keys in hippocampus, values in cortex"],"cache_read_input_tokens":22784,"weakest_assumption_plain":"The load-bearing premise is that the hippocampus stores keys (addresses) while the neocortex stores values (content); this anatomical mapping is inferred from correlational evidence such as hippocampal-dependent reinstatement and optogenetic engram studies, and the paper does not directly show that hippocampal representations act purely as addresses rather than as part of the remembered content.","fun_headline_variants_meta":{"raw":{"variants":["Brain uses key-value memory: hippocampus addresses, cortex content","Forgetting is failed retrieval, not erased memory","Hippocampus stores keys, neocortex stores content","Key-value memory explains why forgetting is a retrieval failure","Brain's key-value store: keys in hippocampus, values in cortex"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000656,"raw_usage":{"total_tokens":2978,"prompt_tokens":897,"completion_tokens":2081,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":2000}},"tokens_in":513,"tokens_out":2081,"duration_ms":16886,"temperature":1.0,"reasoning_tokens":2000,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:59:14.906505+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a study found that a specific memory could be behaviorally recalled in full detail while its hippocampal engram was optogenetically silenced—or, conversely, that stimulating a hippocampal engram alone reproduced the memory's content without any neocortical reactivation—the proposed division of labor between hippocampal keys and neocortical values would be contradicted.","supporting_citations":[{"cited_title":"The hippocampal memory indexing theory","cited_arxiv_id":null,"evidence_quote":"Proposes hippocampal memory indexing theory, the direct precursor of the claim that the hippocampus stores keys."},{"cited_title":"Engram cells retain memory under retrograde amnesia","cited_arxiv_id":null,"evidence_quote":"Demonstrates that engram cells persist and can be optogenetically reactivated after retrograde amnesia, supporting retrieval failure over erasure."},{"cited_title":"Changes in context-specificity during memory reconsolidation: selective effects of hip- pocampal lesions","cited_arxiv_id":null,"evidence_quote":"Shows that a reminder restores context-specificity in normal rats but not hippocampal-lesioned rats, evidence that the hippocampus stores keys."},{"cited_title":"Overlap among spatial memories triggers repulsion of hippocampal representations","cited_arxiv_id":null,"evidence_quote":"Reports repulsion of overlapping hippocampal representations during learning, evidence that hippocampal keys are optimized for discriminability."},{"cited_title":"Dissociating memory accessibility and precision in forgetting","cited_arxiv_id":null,"evidence_quote":"Dissociates memory accessibility from precision over retention, showing forgetting reflects loss of access rather than degradation of stored content."}],"review_version":1}