Experimental Associative Memory Projection

This branch explores whether a neuron-astrocyte-inspired associative projection can help Zaxy recover latent agent state from partial event-history cues.

The experiment is based on:

Kozachkov, Slotine, and Krotov, Neuron-Astrocyte Associative Memory: <https://arxiv.org/abs/2311.08135>
The authors' NAAM reference repository: <https://github.com/kozleo/naam>

Scope

This is not a production checkout feature and does not change Zaxy core. Eventloom remains the immutable source of truth. The associative projection is a derived, replayable, discardable layer that can only propose candidates after resolving them back to cited Eventloom event refs.

The target is not LongMemEval answer synthesis. Zaxy already has strong LongMemEval-compatible results. The target is partial-cue pattern completion:

recover a hidden root cause from scattered symptoms;
recover an unstated coordination constraint from indirect handoff failures;
surface distributed state that is not directly named in the user's query;
preserve citations for every recalled pattern.

Current Baseline

The first implementation is intentionally small and dependency-free:

src/zaxy/associative_memory.py
zaxy experimental pattern-completion
tests/test_associative_memory.py

It uses a deterministic token projection:

Tokenize Eventloom events into event packets.
Seed active process terms from direct lexical matches.
Reinforce terms through shared event packets over a few iterations.
Rank final candidate events by direct and associative support.
Score candidates only after Eventloom refs and hashes are present.

This approximates the useful software lesson from the paper without importing the research model directly: a slower higher-order process layer can propagate weak cues across multiple local events.

Run

zaxy experimental pattern-completion --output-dir reports/benchmarks/pattern-completion-smoke

For machine-readable output:

zaxy experimental pattern-completion --output-dir reports/benchmarks/pattern-completion-smoke --json

The command writes:

pattern-completion-workload.json
pattern-completion-benchmark.json
pattern-completion-benchmark.md
one Eventloom JSONL file per case

StateRecoveryBench

StateRecoveryBench is the broader falsification harness for this branch. It uses adversarial event histories for hidden causes, release constraints, stale runtime state, release-fixture drift, user quality bars, coordination metric gaps, generated authority-shaped negatives, incomplete authority metadata, bridge-evidence requirements, and no-safe-answer abstention cases.

Run it with:

zaxy experimental state-recovery --output-dir reports/benchmarks/state-recovery-smoke

For machine-readable output:

zaxy experimental state-recovery --output-dir reports/benchmarks/state-recovery-smoke --json

The built-in baselines are:

direct_lexical: direct query-token overlap.
hash_vector: deterministic hashed token-vector cosine similarity.
graph_traversal: lexical seed plus shared-term traversal.
zaxy_core_proxy: deterministic source-aware proxy for current Zaxy retrieval posture; this is not a full MemoryFabric checkout run.
memory_fabric_checkout: appends each case through MemoryFabric, runs the model-facing Memory Checkout contract with the coordinate purpose profile, and scores selected cited checkout facts/evidence.
associative_projection: the derived process-term propagation baseline.
authority_resolved_associative: associative projection followed by explicit current/promoted/non-rejected authority filtering.

Current official StateRecoveryBench scores over the expanded 33-case workload are tracked in reports/benchmarks/state-recovery-v1/state-recovery-benchmark.md:

baseline	state accuracy	minimal evidence recall	stale rejection	distractor resistance	abstention accuracy	token cost	latency ms	citation coverage
associative_projection	1.000	0.803	0.485	0.212	0.848	56	0.328	1.000
authority_resolved_associative	1.000	0.985	1.000	0.909	1.000	28	0.348	1.000
direct_lexical	0.697	0.606	0.394	0.121	0.848	69	0.049	1.000
graph_traversal	0.818	0.742	0.515	0.061	0.848	71	0.064	1.000
hash_vector	0.697	0.606	0.394	0.121	0.848	69	0.037	1.000
memory_fabric_checkout	0.818	0.909	1.000	0.818	1.000	34	436.638	1.000
zaxy_core_proxy	0.697	0.652	0.424	0.091	0.848	71	0.045	1.000

The first StateRecoveryBench run showed that associative propagation recovered latent state but pulled in nearby symptom/distractor observations. The expanded workload makes that failure mode explicit: raw association is strong at latent state recovery, but weak at authority, stale-row rejection, and abstention.

The authority-resolved baseline keeps broad propagation for recall, then resolves from the associative support set using explicit current/promoted/non- rejected authority metadata and current observation bridge rows. This preserves latent state recall, improves minimal evidence recall, rejects stale rows, handles no-safe-answer cases, and cuts the average token budget roughly in half. It is still branch-local research code, not a production checkout policy.

The MemoryFabric checkout-backed run is the important product signal. It is not the proxy: every case is appended through MemoryFabric, projected through the embedded graph backend, assembled through Memory Checkout, filtered by the coordinate purpose profile, and scored only from selected cited checkout facts/evidence. That full path clears the 1.1 production guardrails for state accuracy, minimal evidence recall, stale rejection, distractor resistance, abstention accuracy, and citation coverage. The guardrail is deliberately scoped to accepted-state recovery under this 33-case workload; it does not replace LongMemEval or CoordinationBench.

This run also found and fixed a real metadata propagation gap. Generic Eventloom rows now preserve authority metadata such as authority_scope, status, stale, promoted, and superseded_by through verbatim and graph checkout lanes, so purpose policies can suppress stale, rejected, unsupported, and worker-local rows consistently.

The non-perfect distractor score is useful. The remaining misses are old hand-authored cases where distractors lack enough authority metadata to reject them, or where a bridge event is not cleanly distinguishable from nearby support. That argues for authority/provenance metadata as the real Zaxy thesis, not for adding a heavier neural associative model just because the direction is interesting.

Metrics

latent_state_recall: whether expected latent-state terms are recovered.
evidence_recall: whether expected support events are retrieved.
citation_coverage: whether returned evidence has Eventloom hashes.
injected_tokens: approximate support-term count.
latency_ms: projection completion latency.

The same harness reports a direct lexical baseline. The experiment only earns attention if the associative projection beats direct lexical retrieval on partial-cue state recovery while keeping citation coverage at 1.0.

StateRecoveryBench adds:

state_accuracy: exact all-terms latent-state recovery.
minimal_evidence_recall: required proof events returned.
stale_rejection: stale/superseded events avoided.
distractor_resistance: distractor events avoided.
abstention_accuracy: no-answer cases return no evidence instead of a plausible but unsupported answer.
token_cost: retrieved/support term budget.
latency_ms: baseline runtime.
citation_coverage: returned evidence has Eventloom hashes.

Next Research Gate

The current branch is a baseline. The next useful gate is not a heavier Modern Hopfield or NAAM-like implementation yet. The next gate is an accepted-state checkout resolver that can choose the promoted state packet from a broad Memory Checkout result while keeping the broader evidence auditable in provenance.

Only if that benchmark shows a remaining class of failures that event-sourced authority/provenance cannot solve should Zaxy consider learned/tensor process state.