Living Memory — decay, strengthening, association

How the hybrid-memory plugin behaves as a living system: memories strengthen when used, decay when they aren’t, bond when recalled together, carry the context they were formed under, and consolidate when they express the same thought. This page documents the mechanics and every config knob added by the living-memory upgrade (all on by default unless marked).

Strengthens when recalled

  • Every genuine full-content recall — ambient injection and explicit memory_recall — bumps recall_count/access_count, renews the decay TTL, and nudges confidence asymptotically toward 0.95 (5% of the remaining gap per recall; 1.0 stays reserved for verified/confirmed). Time-travel reads (asOf) never count.
  • Hebbian bonding (graph.strengthenOnRecall, default true): facts recalled together strengthen their RELATED_TO links (+0.1, top-8 co-recalled per recall). Recall shapes the graph.
  • Recall also extends a fact’s decay half-life (up to 3× — see below), so used memories age slower, not just later.

Decays — as a curve, not a cliff

  • Confidence half-life (maintenance.decay.mode, default "half-life"): confidence decays continuously per content-type half-life (decision/preference/edict: never; handoff 30d; conversation/progress 45d; fact/note 60d; research 90d; project 120d), extended by recall. Facts falling below 0.1 confidence are deleted. "cliff" restores the legacy one-shot halving as an escape hatch. maintenance.decay.secondChance (default true) gives important (≥0.7) or frequently-recalled (≥3) facts exactly one TTL/2 reprieve at expiry instead of deletion.
  • Link decay (graph.linkDecay { enabled: true, halfLifeDays: 30, floor: 0.05 }): unused RELATED_TO links lose half their strength per half-life and are pruned below the floor — the counterpart that keeps Hebbian bonding from saturating every edge at 1.0. Typed/curated links (PART_OF, CAUSED_BY, …) never decay.
  • Pinned facts are exempt from all decay deletion, and durable now means 180d (was 90d, indistinguishable from normal).

Neighbors at formation

  • Universal auto-linking (graph.autoLink, now default true; budget graph.autoLinkBudgetPerMin: 30): every newly stored fact — auto-capture, distill, reflection, consolidation, CLI, GraphQL, not just the memory_store tool — gets semantic + entity links asynchronously at formation. A fact without edges can never be found associatively.
  • Edges across time (graph.temporalEdges, default true): consecutive facts of a session are chained with PRECEDED_BY links (strength 0.3), giving recall and the Memory Graph a temporal trail. Temporal edges neither strengthen nor decay.

Associative recall

  • Auto-recall graph expansion (graphRetrieval.autoRecallExpand { enabled: true, maxAdds: 5 }): after the ambient pipeline ranks its candidates, 1-hop neighbors of the top-3 seeds join the candidate set with hop-decayed scores — injected context includes what the memory associates with the topic, not just what embeds like the prompt. Association follows strong meaning-edges only: PRECEDED_BY hops and links below strength 0.4 are excluded (session adjacency is not relevance).
  • Explicit memory_recall keeps its existing GraphRAG expansion; co-activation ranking (facts history recalls as a group boost each other) now computes from real recall_events co-occurrence when composite-score v2 is enabled.
  • Serendipity slot (autoRecall.serendipity, default true; { cooldownPrompts: 10, minLinkStrength: 0.4, staleImportanceMin: 0.7, staleDays: 30 }): once per cooldownPrompts prompts, one labeled [serendipity] headline joins ambient injection — weighted-random from strong-but-never-recalled graph neighbors of the current results, falling back to stale-important facts. Index-only exposure: a 60-char title, never full text, recall_count untouched — a memory that surfaces this way stays eligible until you actually ask about it.

Same-thought consolidation

  • Consolidation clusters by embedding cosine and by claim: facts with the same non-empty (entity, key) in the same scope merge even when phrasing lands below the 0.92 cosine threshold. Cadence raised from ~monthly to every 5 days (consolidate step) so similar facts merge before their TTL wins.

Emotional state & routines

  • Affect stamping (nightly affect-stamp step, active when frustrationDetection.enabled): the frustration detector’s per-session signals stamp a confidence-weighted valence (−1..1) and affect_source onto facts formed in that session’s window. Memory formation carries emotional state; nothing runs on the hot path.
  • Routine mining (maintenance.routineMining { enabled: true, maxPerRun: 2, timezone: "UTC" }, nightly): recall patterns recurring ≥3 times across ≥3 distinct weeks in the same weekday/time-band become ordinary decayable routine facts (“Routine: on Tuesday mornings, a recurring focus is …”). Weekday/time-band bucketing uses timezone (an IANA zone, default "UTC") — set it to the user’s local timezone so a “night” routine means local late-night, not 00:00-06:00 UTC. Routines that stop recurring decay out — learned from interaction, not programmed.

Free-text contradictions (shipped)

  • Contradiction candidates (maintenance.contradictions { freeText: true, similarityFloor: 0.85, maxPairsPerRun: 40, minConfidence: 0.7 }, nightly): the structured detector only sees exact entity+key collisions — this pass closes the free-text gap. Recent facts (48h) → vector top-k in-scope neighbors at cosine ≥0.85 → NLI verdict (nano tier, temperature 0) → recordContradiction with the nli_free_text audit marker and a CONTRADICTS link, flowing into the same nightly resolve pass and the Memory Graph conflicts panel. Near-duplicates (consolidation’s job) and same-entity+key pairs (the structured detector’s job) are excluded.

Staged (off by default — measured)

  • Composite-score v2 + MMR diversity (retrieval.compositeScore.v: 2; retrieval.diversity { enabled, mode: "mmr", mmrLambda: 0.7 } — mode/mmrLambda now actually parse): the flip criterion (arm B ≥ arm A on both nDCG@10 and P@5) was measured and failed on the gold set — armA(v1) nDCG@10=1.000 · armB(v2+mmr) nDCG@10=0.996, equal P@5 — so both stay opt-in. Caveats recorded honestly: the fts5 fixture’s ranking is already saturated (any reorder can only cost) and applyMMR runs on its bigram fallback until candidate vectors are threaded through. tests/retrieval-ab-composite.test.ts re-measures on every CI run and fails if v2 ever regresses nDCG by >0.02; a richer gold set with real embeddings is the path to a flip.
  • Session-start briefing (autoRecall.retrievalDirectives.sessionStart): when enabled, the briefing also resurfaces up to 3 stale-important memories (importance ≥0.7, untouched 30+ days). (Overnight research briefings deliver independently of this flag — see PROACTIVE-RESEARCH.md.)

Measurement

benchmark/retrieval-eval/ + tests/retrieval-eval-harness.test.ts pin retrieval quality (P@5 / R@10 / nDCG@10 over a deterministic 200-fact fixture, plus an ambient-injection hit-rate sample) as CI floors — ranking regressions fail CI. tests/retrieval-ab-composite.test.ts keeps the v1-vs-v2+MMR comparison honest on every run. Decay behavior is pinned by fast-forward survival tests in tests/living-memory-dynamics.test.ts (curve composition, pinned survival, one-shot second chances, recall-extended half-lives).

Built on top of this: the proactive research loop

The observation layer documented here (valence, routines, frustration signals, patterns) now feeds an initiative loop — nightly insight synthesis → deterministic trigger → overnight web research by a cron agent → morning briefing. See PROACTIVE-RESEARCH.md.

Back to top

OpenClaw Hybrid Memory — durable agent memory

This site uses Just the Docs, a documentation theme for Jekyll.