feat: brain-fidelity recall substrate (PPR + Gaps 1/3/4/5)#39
Merged
Conversation
The recall path already computed per-seed activation weights from vector relevance (core stageActivate: seedActivations from m.relevance, plus entity and project seeds) and passed them through the graph port, but the final layer dropped them. NeuralGraph.spreadActivation forwarded only seedNodeIds, and SpreadingActivation.activate hardcoded `reduce(activation = 1.0, ...)`. The graph therefore ran a uniform exponential-decay walk instead of personalized (seed-weighted) spreading activation — the PPR mechanism it was supposed to provide was unplugged one layer above the Cypher. Forward opts.seedActivations into activate() and bind it as the walk's initial value: `reduce(activation = coalesce($seedWeights[seedId], 1.0), ...)`. Seeds with no weight default to 1.0, so callers that pass no map are byte-identical to before; only callers already supplying weights (recall, pattern-completion) change behavior. Tests: unit (stub driver) asserts the weights reach the query params and the Cypher binds them — the exact regression — and runs anywhere; an integration test asserts a 9:1 seed-weight ratio yields a 9:1 activation ratio with a 1:1 unweighted control (requires NEO4J_TEST_READY).
Brain-fidelity recall-quality work (Gaps 3 + 1 from the architecture spec).
Both attack the move-zero finding that the associative layer is poisoned by
noise and duplicate episodes.
Gap 3 - salience-gated plasticity: scoreSalience already scored every episode
(acknowledgment 0.10, decision 0.90) but association edges used flat constants,
so a heartbeat and a decision wired equally strong. New salienceGate() modulates
edge strength by the WEAKEST-LINK salience of the two endpoints (a salient
partner cannot rescue a noise endpoint) and drops the edge below a low-cut
("no plasticity without salience"). Wired into createTemporalEdges using the
salience already in-hand at ingest. Backward-compatible: no salience map keeps
the flat 0.3 behavior byte-identical.
Gap 1 - pattern separation (pragmatic tier): ingest inserted unconditionally,
so repeated turns (heartbeats, re-emitted tool output) piled up as distinct
rows and interfered at recall. New findNearDuplicate() checks the incoming
embedding against recent same-session episodes; on a hit (cosine >=
dedupeThreshold, default 0.95) the existing memory is reinforced via
recordAccess instead of storing a redundant copy. Conservative, configurable
(MemoryOptions.dedupeThreshold; <=0 disables), and non-destructive.
Pure primitives (salienceGate, findNearDuplicate) are unit-tested; the ingest
merge + reinforcement is covered end-to-end. core 509/509, tsc clean.
Deferred (storage boundary): gating co-recalled edge strength
(upsertCoRecalled) and dream-cycle replay/topical weights; the k-WTA
separation code (Gap 1 tier 2).
Extends salience-gated plasticity to the two remaining frequency-weighted edge sites — the Hebbian co-recall layer that move-zero identified as the poisoned one (heartbeats co-recalled repeatedly, accumulating high strength), and the dream-cycle hippocampal-replay edges. Rather than thread a strength delta through the upsertCoRecalled storage signature (which on PostgREST means an engram_upsert_co_recalled RPC + a schema migration on prod), the gate is applied at the edge-creation DECISION: a co-recall edge whose least-salient endpoint is noise is simply not created or strengthened. Same un-poisoning, zero schema/deploy risk, identical on both storage adapters. - createCoRecalledEdges: accepts optional per-memory salience and skips a pair when salienceGate() drops it (weakest-link; unknown salience -> 0.5, ungated). - reconsolidation: looks up encoding salience for the co-recalled episodes and passes it through. Non-episode (curated semantic/procedural) stay ungated. - dream-cycle replay: the seed query returns m.salience and the TOPICAL replay edge weight is salience-gated, so a replay edge between noise seeds forms weakly or not at all. Backward compatible: every existing caller passes no salience -> 0.5 -> unchanged. core 514/514, tsc clean. Deferred: SQL discoverTopicalEdges + causal/light/deep-sleep weights (lower, diffuse value); the graded strength-delta refinement via the RPC.
…s 4 + 5) Both wire signals that were already computed but never used at retrieval. Gap 4 - context reinstatement: the sensory buffer already tracks the topics primed by recent turns (the active conversational context), but recall seeded spreading activation only from query relevance + entities + project. Now the primed topics are looked up and added as soft context seeds (weight 0.45, below entity/project seeds), so the SAME query completes to different associations depending on what is currently in focus (encoding specificity). Best-effort and query-dominant. Gap 5 - lateral inhibition / fan effect: dream-cycle's betweenness pass already flags high-centrality hubs with isBridge, but recall ignored it. Activated nodes are now down-weighted (x0.5) when isBridge, applied before the primary/faint threshold split so a generic hub can fall out of the result set entirely. No-op when betweenness was never computed (no isBridge property present). Both land in stageActivate; the engine passes sensory.getPrimed() through. Backward compatible (no context topics -> unchanged; no isBridge -> unchanged). 4 new tests, core 518/518, tsc clean. Deferred: degree-based divisive normalization in the activation Cypher (a heavier, always-available alternative to the isBridge gate); per-topic boost weighting of context seeds.
|
Caution Review failedPull request was closed or merged during review WalkthroughAdds salience-gated plasticity and near-duplicate ingestion, gates temporal/co-recall/topical edges by endpoint salience, enables per-seed personalized spreading activation, and reinstates sensory context and hub inhibition during recall. ChangesSalience-gated memory associations and semantic deduplication
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
packages/core/src/consolidation/dream-cycle.ts (1)
463-476:⚠️ Potential issue | 🟡 MinorAlign the missing-salience fallback in dream-cycle with salienceGate semantics.
Non-episode Neo4j
:Memorynodes (e.g., semantic/procedural) are created withoutsalience, so the 0.5 fallback is expected—butsalienceGateonly drops edges whenmin(salienceA, salienceB) <= 0.15(DEFAULT_LOW_CUT). Since episode DB default is 0.3, both 0.3 and 0.5 bypass “gated out”; 0.5 mainly boosts TOPICAL edge weights (0.5 => full baseStrength, 0.3 => ~0.6×). Switch the fallback to 0.3 or update the comment to describe this weighting effect, not “never gated out.”🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@packages/core/src/consolidation/dream-cycle.ts` around lines 463 - 476, The seeds construction in dream-cycle uses a 0.5 fallback for missing m.salience which wrongly claims it "never gated out" and skews TOPICAL weighting; change the fallback to 0.3 to match episode DB default and salienceGate semantics (DEFAULT_LOW_CUT) or alternatively update the comment to explain that 0.5 increases TOPICAL edge weights rather than preventing gating; modify the code that builds seeds (the seedResult.map → seeds block) to use 0.3 as the default salience value (or adjust the inline comment) and ensure this aligns with salienceGate/DEFAULT_LOW_CUT behavior referenced elsewhere.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Outside diff comments:
In `@packages/core/src/consolidation/dream-cycle.ts`:
- Around line 463-476: The seeds construction in dream-cycle uses a 0.5 fallback
for missing m.salience which wrongly claims it "never gated out" and skews
TOPICAL weighting; change the fallback to 0.3 to match episode DB default and
salienceGate semantics (DEFAULT_LOW_CUT) or alternatively update the comment to
explain that 0.5 increases TOPICAL edge weights rather than preventing gating;
modify the code that builds seeds (the seedResult.map → seeds block) to use 0.3
as the default salience value (or adjust the inline comment) and ensure this
aligns with salienceGate/DEFAULT_LOW_CUT behavior referenced elsewhere.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: ASSERTIVE
Plan: Pro
Run ID: 7b4b2120-f995-4077-a9c7-cdb1cc76f283
📒 Files selected for processing (17)
packages/core/src/consolidation/dream-cycle.tspackages/core/src/ingestion/near-duplicate.tspackages/core/src/ingestion/plasticity.tspackages/core/src/memory.tspackages/core/src/retrieval/engine.tspackages/core/src/retrieval/reconsolidation.tspackages/core/src/retrieval/spreading-activation.tspackages/core/src/systems/association-manager.tspackages/core/test/ingestion/near-duplicate.test.tspackages/core/test/ingestion/plasticity.test.tspackages/core/test/memory.test.tspackages/core/test/retrieval/spreading-activation-gaps.test.tspackages/core/test/systems/association-manager.test.tspackages/graph/src/neural-graph.tspackages/graph/src/spreading-activation.tspackages/graph/test/spreading-activation-ppr.test.tspackages/graph/test/spreading-activation-seed-binding.test.ts
The Gap 1 near-dup merge defaulted to 0.95-on, which collapsed the artificially near-identical embeddings used by e2e isolation tests (shared-axis fake embedder) and broke CI in @engram-mem/sqlite. More fundamentally, the merge's recall benefit is unvalidated — by the same gate that kept pattern completion opt-in, dedup should be opt-in until measured on real data. dedupeThreshold now defaults to 0 (off); set ~0.95 to enable. No ingest behavior change by default. The Gap 1 unit test opts in explicitly. sqlite 106/106, postgrest 71/71, core 518/518.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Production recall-quality changes from the brain-fidelity architecture work — a
brain-model memory that reconstructs the correct context from a partial cue,
reliably. Each commit wires a mechanism the architecture had the organ for but
was not using.
Commits
activation = 1.0, dropping the per-seed vector-relevance weights that corealready computed and passed through the port. Now bound:
reduce(activation = coalesce($seedWeights[seedId], 1.0), ...).near-identical recent turns reinforce the existing memory instead of storing a
redundant copy (pattern separation); temporal edges are gated by the salience
already scored at ingest (weakest-link).
Hebbian co-recall layer (the poisoned one) and dream-cycle replay edges no
longer form between noise endpoints. Gated at the edge-creation decision, so no
storage schema/RPC change.
seeds from the active conversational context (primed topics), and
high-betweenness hub nodes (isBridge) are suppressed at recall.
Compatibility
Additive and backward-compatible: no salience map -> unchanged edge weights;
dedupeThreshold: 0-> unchanged ingest; no primed topics / no isBridge ->unchanged recall. PPR affects only the associations channel, not primary ranking.
Test plan