연구 기록 사본

원 저장소 · 원본 neural-memory/RESEARCH.md · 수집 2026-10-16 · 커밋 fd190b18f8e4113d7f451a7108911c1f89716b24 당시의 실험 조건·수정·철회를 보존한 문서입니다. 성능 수치를 서로 다른 실험 간에 합산하지 않습니다. 연구 안내와 해석 범위부터 읽으세요. 코드·데이터·체크포인트는 이 사이트에 포함하지 않았습니다.

Research reuse map

This experiment deliberately combines established mechanisms instead of proposing a new architecture before the basic phenomenon is measured.

Prior work	Mechanism reused here	Deferred from phase 0
Learning to (Learn at Test Time)	A model-valued hidden state updated online; slow parameters learn the update behavior	Full TTT language-model stack and optimized mini-batch scan
Titans	Prediction residual as surprise; explicit retention/decay	Momentum, deep memory MLP, and integration into a transformer backbone
Linear Transformers Are Secretly Fast Weight Programmers	Residual delta-rule update, which can replace an existing key/value association	Kernelized multi-head fast weights
MIRAS	Separate the choices of memory architecture, attentional objective, retention, and optimizer	Alternative objectives and retention rules
Fast Weight Memory	Evaluate compositional associative behavior as a learned recurrent state	Its LSTM controller and task-specific reasoning stack
Synaptic plasticity-dependent competition rule influences memory formation	Treat memory allocation as competition among plastic parameter groups, not only a global write/no-write decision	Any claim that mouse fear-memory allocation is a general model of human memory

The official TTT PyTorch implementation is a useful equation and tensor-shape reference, but its authors describe it as a slow tutorial implementation and do not recommend it for training. Copying the full language-model stack would therefore increase the first experiment’s cost without testing the write-selection hypothesis.

Frozen design choices for phase 0

• Memory representation: one matrix-valued fast-weight model state per episode.
• Write rule: normalized residual delta update.
• Attentional objective: later key-to-value recall cross-entropy.
• Retention: fixed multiplicative decay.
• Write selection: learned scalar gate conditioned on the observation, noisy present-time evidence, and prediction surprise.
• Outer supervision: future recall only. retained is used for episode construction and evaluation, never as a direct write-gate label.
• No graph or external store: the changing parameter tensor is the memory.

These choices isolate one question: can delayed future utility teach a model when to change its own fast weights? They do not yet establish semantic text memory.

Decision gates

1. Mechanism gate: learned write gating must beat an identically trained always-write model on held-out random bindings, while assigning higher mean gates to useful events.
2. Dynamics gate: add correction, delayed decay, and multi-timescale consolidation only after gate 1 passes.
3. Language gate: replace symbolic IDs with frozen text encoder representations and add a small decoder only after the state-update mechanism passes.
4. Agent gate: connect the memory output to a frozen LLM and measure spontaneous recall without memory instructions only after the standalone model passes.

Moving in this order prevents a capable language decoder from hiding a failed memory mechanism.

Biological allocation hypothesis for phase 1

Jeong et al. manipulated potentiation and depression at auditory inputs to the lateral amygdala in mice. Potentiated inputs were preferentially recruited into the fear-memory engram; post-training depression could reverse that allocation. The total recall-active population remained approximately constant, which the authors interpret as evidence for competitive allocation under limited plasticity rather than independent binary selection by every neuron.

This suggests a specific computational ablation, not a biological equivalence claim:

single global write gate              competitive block allocation
g = sigmoid(controller(x))     vs.    a = sparse competition(score_1 ... score_n)
W <- W + g * delta                    W_i <- W_i + a_i * delta_i

The competition should have a fixed plasticity budget (sum(a) = 1 or top-k allocation). A recent event can still be reassigned during a short eligibility window: strengthening one block competes with or depresses neighboring blocks. The testable question is whether this improves capacity, correction, and memory linking over the scalar-gate phase-0 model at an equal parameter and update budget.

Do not use this study as evidence that the same rule governs all human memory. It examined a specific fear-conditioning circuit in mice with artificial optogenetic LTP/LTD.

Survey, 2026-09-14: should the write decision be deferred?

Phases 12 to 16 train a write gate to predict future utility from the present observation. Phase 7 already measured the ceiling of that framing on real logs: 0.543 balanced accuracy from observation-time evidence against 0.980 once delayed feedback arrives. This survey asks whether the deferred alternative has support beyond that single measurement, and what shape it should take.

The boundary from phase 1 still applies, but it applies to one use only. Animal results are a source of architecture to test, never evidence for a claim about human memory or about our own numbers. Phase 1 is the precedent: competitive allocation came from Jeong et al. and beat a single matrix 0.9243 to 0.2347, which is a result about our model, not about mice.

Deferred consolidation in biology

Three independent mechanisms converge on writing cheaply from a local signal and letting a later event decide what survives. What they add beyond the principle is a time constant.

Mechanism	Deferral window	Source
Excitability-biased allocation (CREB); excitability decays over hours	hours	Josselyn & Frankland 2015; J Neurosci 2024
Synaptic tagging and capture: a weak tag captures later plasticity-related proteins	classically under 3 h; one 2026 report extends capture to 9 h	Frey & Morris, Nature 1997; Commun Biol 2026
Replay selection biased by reward prediction, consolidated across sleep	hours to overnight	Nat Commun 2026; eLife 2015
Reconsolidation: retrieval returns a trace to a labile state	about 6 h	Nader et al., Nature 2000

The design consequence is that the deferral window is finite. “Hold every candidate until the query arrives” is retrieval, not memory, and has no biological counterpart. The measurable quantity is how far a bounded delay closes the gap between deciding at write time and deciding with the query visible, which on the Claude corpus is 0.5686 against 0.8206 top-1.

The mushroom body as a specified circuit

The fly connectome makes a complete memory circuit available as architecture rather than principle: 139,255 neurons and about 50 million synapses in the adult female brain (Dorkenwald et al., Nature 2024; Schlegel et al. 2024). Connectivity alone, with neurotransmitter identity, already predicts behaviour in a whole-brain leaky integrate-and-fire model (Shiu et al., Nature 2024).

Four properties bear on our design.

• Sparse expansion. About 50 projection neurons fan out to roughly 2,000 Kenyon cells, each integrating 6 to 8 inputs across 5 or 6 claws. Reciprocal inhibition through the single GABAergic APL neuron keeps only 5 to 10 percent of Kenyon cells active, and removing that sparsening impairs discrimination of similar but not dissimilar odours (Lin et al., Nat Neurosci 2014). Our keep_ratio of 0.25 is several times denser than this.
• The write signal does not read the content. Kenyon cells carry odour identity; dopaminergic neurons carry reinforcement and gate plasticity at the Kenyon-cell-to-output-neuron synapse. The same dopaminergic neuron writes aversive memory when it fires within about 30 s after odour onset and appetitive memory when it precedes odour by 20 to 60 s, so sign comes from the modulatory channel alone. A second activation minutes after training, with no odour present, nearly abolishes the conditioned response (Aso & Rubin, eLife 2016). SelectiveWriteRecallMemory does the opposite: its gate scores the content itself. Phase 7’s apply_outcome() is the component that already matches this shape and is currently unused by the phase 12 to 16 line.
• Compartments are parallel, not a promotion pipeline. Output neurons tile the lobes into 15 compartments with their own dopaminergic input and their own decay, capacity, and flexibility. γ1pedc holds strong immediate memory that is largely gone by 24 h and has a capacity of one, retaining only the most recently trained odour, while α1 has weak immediate memory, supports long-term consolidation, and holds at least two. Trained with opposing signals, both traces form simultaneously and the expressed valence flips as the fast compartment decays. This favours several stores with different decay rates read as a sum, over a provisional buffer that promotes into a durable one.
• Derived algorithms exist but not for this. Fly-inspired locality-sensitive hashing (Dasgupta, Stevens & Navlakha, Science 2017), Bloom-filter novelty detection (PNAS 2018), and sparse coding against catastrophic forgetting (arXiv:2107.07617) all come from this circuit. None of them learns a write gate.

The reinforcement side of the same circuit

Surveyed 2026-09-14, after the modulatory pathway failed twice as a single scalar gain.

The input side is genuinely multi-channel. The hemibrain resolves 21 dopaminergic types, 6 PPL1 and 15 PAM, tiling the 15 compartments roughly one type each, about 130 neurons in total (Li et al., eLife 2020). The dimensions are qualitatively distinct rather than one axis at different gains: PPL1-γ1pedc reinforces shock, heat and bitter while separately carrying hunger-state gating, and PAM-γ5 reinforces sugar and also signals shock omission to extinguish an aversive memory, splitting into five subtypes by input connectivity (Otto et al., Curr Biol 2020). Tuning, sign and time constant vary independently, and the same DAN writes opposite signs depending on its timing against the odour.

The output side is opponent and summed. Twenty-two canonical output types plus fourteen atypical, in three transmitter classes split by lobe. An output neuron’s valence is opposite to its compartment’s dopaminergic valence, and the plasticity is depression: the reinforcement turns down the output that signals the opposite valence rather than strengthening one that agrees (Owald et al. 2015; Aso et al., eLife 2014). Behaviour follows a weighted sum across compartments that is balanced in a naive animal, not a winner.

The output feeds back into the signal that writes it. The connectome shows extensive and selective output-to-dopaminergic feedback onto DAN dendrites: MBON-γ5β′2a contacts PAM-γ5(fb) exclusively, while MBON-γ4γ5 is the strongest input to PPL1-γ1pedc and does not touch PAM-γ5 (Li et al. 2020; Otto et al. 2020). Modelling reads that return path as reinforcement prediction error computed by subtraction across the opponent pair (Bennett, Philippides & Nowotny, Nat Commun 2021).

Two channels reproduce the behaviour. Bennett et al. collapse the fifteen compartments to an approach and an avoidance channel and recover a broad range of conditioning and blocking results. The two differ in sign alone, not in tuning or time constant.

What this says about our failures

Phase 18 and the associative attempts both ended with a modulatory gain saturating at its bound and contributing nothing. Two circuit properties we had omitted explain that directly. A gain with no prediction subtracted has no reason to stop growing, whereas a prediction-error drive falls to zero once the memory already answers the event. And a single positive channel can only rescale the state, which is rank-preserving and therefore provably unable to reorder what is retrieved; an opponent pair read as a difference expresses “strengthen this, weaken that”, which is what competitive allocation needs.

The licensed change is two signed channels sharing one eligibility trace, with the readout as their difference and the drive computed as reinforcement minus what the memory already predicts. Per-channel decay is the better second step than more channels, since Phase 22 showed tag decay is the parameter that creates the rescue window.

Contested or unverified. Whether one dopaminergic type per compartment is universal; DAN-to-DAN connectivity is absent from both sources, which is absence of evidence rather than a demonstrated negative; and the two-channel sufficiency is a modelling claim about conditioning assays, not about memory capacity or longevity.

What forgetting research says about capacity

Forgetting is an active, default-on process rather than passive leakage: ongoing dopaminergic activity erases labile memory in Drosophila unless salience marks it for retention (Berry et al., Neuron 2012), and transience is argued to be functional, preventing overfitting to specific episodes (Richards & Frankland, Neuron 2017). Decay rate itself appears to adapt to update frequency, existing to limit interference (Altmann & Gray, Psych Science 2002).

Competition bounds relative excitability, not a counted slot budget, and Landauer explicitly disclaimed his capacity estimate as a bound on storage (Landauer, Cognitive Science 1986). A hard top-k at write time therefore collapses three separable knobs — encoding strength, decay rate, and retrieval threshold — into one irreversible switch. Phase 14 already screened an input-dependent retention gate and rejected it, attributing the failure to 775 training episodes being too few to identify a richer retention policy rather than to the formulation.

Where our novelty now sits

Prior work	Relation to this project
Learning to Evict from Key-Value Cache (KVP, ICML 2026)	Per-head reinforcement-learned agents rank tokens by predicted usefulness. This is already a learned future-utility write gate, so “learned write gating” is no longer a differentiator. Differs by acting on KV tokens inside the model rather than semantic traces beside a frozen encoder, and by being single-stage.
RecMem (ACL 2026 Findings)	Closest two-stage system: every interaction enters a subconscious embedding layer and an LLM consolidates only when sustained recurrence appears among semantically similar interactions, cutting construction tokens up to 87%. Promotion-only and never discards, so it does not face irreversible capacity loss.
What Eviction Destroys	Independent support rather than competition: a restore-counterfactual audit attributing 60 to 73 percent of corrected errors to irreversible eviction at an 80k budget, and all of them at 8k. Corroborates the phase 15 finding that the whole task-level shortfall is write selection.
Agentic Memory (ACL 2026)	Learns unified long- and short-term memory management with three-stage reinforcement learning and a step-wise GRPO variant for sparse memory-action rewards. Whether the backbone language model is itself updated is not stated in the abstract and was not verified.

The remaining differentiators are the two-stage deferral and the irreversible-discard semantic-trace setting, not learned write gating as such.

Contested — do not present as settled

• Post-encoding reversal of allocation rests mainly on Jeong et al. 2021. General allocation reviews do not establish it.
• Human retroactive behavioural tagging has explicit null results, and reconsolidation has a direct failed replication (Sci Rep 2022).
• Parallel compartments versus the older sequential consolidation model (γ to α′β′ to αβ) remains open; trace migration must not be presented as fact.
• Whether forgetting erases or only removes access is disputed (Ryan & Frankland, Nat Rev Neurosci 2022).
• Kenyon cell connectivity is mostly but not entirely random; food-responsive projection neurons are a reported exception (eLife 2022).
• Neuromem is a streaming-memory testbed that decomposes the lifecycle into data structure, normalization, consolidation policy, query formulation, and context integration. It treats consolidation as a distinct dimension, but whether any stage carries learned parameters is not stated in the abstract and was not verified.

All arXiv identifiers in this survey were checked against the listing pages on 2026-09-14.

Redesign, 2026-09-14

Why the previous design has to be abandoned

Phases 12 to 19 asked a bounded layer to pick which of eight candidates a later query would need. Every result in that line collapsed to the same thing once controlled. On the deferred corpus an untrained random projection of the same width already reaches 0.4033 top-1 with no matrix and no learning; the trained associative model reaches 0.4956; a blank event scores 0.4956 against 0.4971 for the correct one; and four different write policies land between 0.4825 and 0.5000. The slot model’s large event effect came from feeding cosine-to-event in as a residual score rather than from anything the memory computed.

The cause is the task shape. Presenting the candidates at recall time lets encoder cosine stand in for memory, and cosine is better at it than any bounded state we can train. Two further points sharpen this. Importance is not utility, so scoring a write decision against a future-utility oracle measures the wrong thing by construction. And memory is gist-preserving and detail-losing, so a metric rewarding exact reconstruction of a stored embedding rewards verbatim storage, which is not what memory does.

None of the survey’s findings are about choosing from a candidate list. Every one of them is about how a bounded memory degrades: transience as a function, decay adapting to interference, compartment time constants, sparsening that helps only for similar items. The experiments must measure that instead.

What the survey licenses, as constraints

1. The write signal must not read content. Dopaminergic neurons carry reinforcement, not odour identity, and sign comes from the modulatory channel alone. Write strength may depend on the memory’s own prediction error or on an external outcome, never on a judgement about the content.
2. Modulation must work with the item absent. A second activation minutes after training, with no odour present, nearly abolishes the response. Consolidation therefore acts on an eligibility trace addressed by coincidence, not by re-presenting the item.
3. Stores run in parallel with different time constants and are read as a sum. Compartments do not promote into one another; γ1pedc holds one item and is gone by 24 h while α1 is weak immediately and durable later, and the expressed answer flips as the fast store decays.
4. Codes are sparse and expanded, and sparsening matters only for similar items. Fifty inputs fan out to two thousand cells with 5 to 10 percent active, and removing the inhibition impairs discrimination of similar but not dissimilar odours.
5. Forgetting is active and its rate adapts to interference. Decay is a trained dynamic, not a constant, and encoding strength, decay rate, and retrieval threshold stay three separate knobs.
6. Allocation is competitive under a roughly fixed budget. Potentiated inputs are preferentially recruited while the recall-active population stays about constant, so strengthening one trace should cost others rather than being free.
7. What is allocated depends on current excitability, which decays over hours. Items encountered close together are therefore biased onto shared substrate, which predicts linking between them and not only interference.
8. A weak trace can be rescued later, inside a window. Tagging and capture give a weak event access to plasticity a later strong event supplies, classically under three hours. The window is finite, which is the part the corpora could not show us.
9. Consolidation continues offline and selectively. Replay is biased by reward and keeps working across sleep, so retention can improve with no new input.
10. Reading a memory changes it. Retrieval returns a trace to a labile state before it restabilises, so recall is not a passive operation on the state.

The replacement task

Store N documents in a bounded state, probe with a cue, and measure what comes back. No candidate list exists at recall, so similarity between the cue and the stored items cannot substitute for the memory. The state is the only thing carried from write to read.

Two scores per probe, because the survey says surface and gist should not be treated alike:

• gist recovery — agreement between what is recalled and the fact the document carried;
• surface recovery — agreement between what is recalled and the document’s own form.

A memory that behaves like the literature describes should lose surface faster than gist as load rises. A verbatim store loses both together; a retriever loses neither until it fails entirely.

Measurements, one per mechanism the survey established

Every row names the finding it comes from, the prediction that follows, and what would falsify it. Elapsed time in the animal work maps onto intervening writes here, since interference rather than the clock is what the state actually experiences.

#	From	Measurement	Prediction	Fails if
1	Transience is functional (Richards & Frankland 2017); decay adapts to interference (Altmann & Gray 2002)	Fidelity against number of documents stored, at fixed state size	Graceful degradation, and a trained decay beats a fixed one only as load rises	Flat, or a cliff, or trained decay never separates from fixed
2	Sparse expansion and APL inhibition; removing it impairs similar but not dissimilar odours (Lin et al. 2014)	The same curve for sparse expanded codes against dense low-dimensional ones, run separately on similar and dissimilar document sets	Sparse degrades more slowly, and the gap appears only for similar documents	Sparsity helps uniformly, or not at all
3	Memory is gist-preserving and detail-losing	Gist against surface recovery across the same load	Surface decays first	Both decay together
4	Compartments run in parallel with their own decay and capacity; expressed valence flips as the fast one fades (Aso & Rubin 2016)	Two stores with different decay read as a sum, probed after varying numbers of intervening writes	What is expressed changes with delay, and one store cannot reproduce the curve	A single store matches the two-store readout
5	Competitive allocation under limited plasticity; total engram population roughly constant (Jeong et al. 2021)	Whether strengthening one trace measurably costs others, at fixed state size	Gains and losses trade off; forcing that trade-off helps rather than hurts	Traces strengthen independently, or the constraint only hurts
6	Excitability biases allocation and decays over hours (Josselyn & Frankland 2015)	Whether documents written close together share substrate	Items written adjacently are recalled together more than distant ones, a linking effect distinct from interference	No difference between adjacent and distant pairs
7	Synaptic tagging and capture: a weak tag is rescued by a later strong event inside a window (Frey & Morris 1997)	Write an item weakly, apply a strong unrelated event after k intervening writes, measure rescue against k	Rescue falls off with k, giving the deferral window the corpora could not supply	Rescue is flat in k, or absent
8	Replay selection is biased by reward and consolidates offline (eLife 2015; Nat Commun 2026)	An offline phase that re-applies stored eligibility with no new input	Selected traces retain better than unselected ones after the same load	Offline replay changes nothing
9	Retrieval returns a trace to a labile state (Nader et al. 2000)	Probe an item, then re-write, and measure drift	Recalled items drift, and repeated recall compounds it	The state is unchanged by being read
10	The dopaminergic write signal never reads content, and a second activation with the item absent abolishes the response (Aso & Rubin 2016)	A modulatory signal applied after writing, with the item absent	Recall changes, and a mismatched signal does not produce that change	The signal is inert, as it was in every attempt today

Order matters. Measurement 2 runs first: it is the sharpest circuit-derived prediction available, it fails in a specific direction, and nothing in this project depends on it yet. Measurements 1 and 3 come with it, since they share the load curve. Measurement 7 is the one worth the most, because it recovers the deferral window that LongMemEval-S could not show — there the later event stayed equally informative at every distance, whereas here the delay is ours to set. Measurement 10 has already failed twice in this project and should not be attempted again until 1 to 3 establish that the memory holds anything at all.

Three of these are marked contested above and must be reported as tests of a disputed claim rather than as confirmations: 5 rests mainly on Jeong et al., 9 has a direct failed replication, and 4 sits against the older sequential-consolidation model.

What has to be built

Controlled document sets rather than found episodes, because measurements 1 and 2 need the number of stored items and the similarity between them as independent variables. LongMemEval sessions supply the documents and their facts; grouping them by encoder similarity supplies the similarity axis. The teacher disappears entirely: the objective is reconstruction of what was stored from its own cue, so no label about which document will matter later enters training at any point.

What this design deliberately gives up

It stops asking whether the layer can predict future utility, which Phase 19 measured as information-limited anyway, and it stops competing with retrieval on retrieval’s own ground. If the memory cannot beat cosine at picking a candidate, that is no longer a finding about the memory. The claim available at the end of this program is narrower and about capacity: how much a bounded state holds, how it fails, and whether it fails the way the biology says it should.