clawmem

内容来源：SKILL.md（标准 Skill 格式） · 原始地址 · 查看安装指南

原始内容

name: clawmem description: "ClawMem operational reference for agents at query time — the 3-rule escalation gate, MCP tool routing, the 4 query-optimization levers, pipeline behavior (query vs intent_search), composite scoring, and memory lifecycle (pin/snooze/forget). Use when tuning retrieval, troubleshooting recall quality, or any ClawMem operation beyond the routing already in your global CLAUDE.md / this repo's AGENTS.md. NOT for setup — install / inference-server config / env vars / systemd / indexing config / internals live in AGENTS.md + docs/." allowed-tools: "mcp__clawmem__*" metadata: author: yoloshii version: 2.0.0

ClawMem Operational Reference

Scope: agent-time operations only — escalation, tool routing, query tuning, pipeline reasoning, composite scoring, lifecycle. Setup, inference-server config, env vars, systemd units, indexing/collection config, graph internals, and the OpenClaw/Hermes plugins are deliberately not here — they live in this repo's AGENTS.md + docs/ (e.g. docs/guides/inference-services.md, docs/reference/configuration.md, docs/troubleshooting.md, docs/internals/). Kept out to avoid drift between this skill and the package.

Routine memory needs neither this skill nor manual MCP calls — hooks + the ClawMem routing already in AGENTS.md / your global CLAUDE.md handle ~90%. Reach for this skill (and Tier-3 tools) only when that isn't enough.

Architecture (one-liner)

Two tiers: hooks = automatic context flow (surfacing, extraction, compaction survival); MCP tools = explicit recall / write / lifecycle. Substrate: QMD retrieval (BM25 + vector + RRF + cross-encoder rerank + query expansion), with SAME (composite scoring), MAGMA (intent + graph), and A-MEM (self-evolving notes) layered on top. Do not call standalone QMD tools.

Tier 2 — Automatic retrieval (hooks)

Hooks handle ~90% of retrieval at zero agent effort.

Hook Trigger Does
context-surfacing UserPromptSubmit retrieval gate → profile-driven hybrid search → FTS supplement → file-aware search → snooze/noise filters → spreading activation → memory-type diversification → tiered injection → <vault-context> (+ optional <vault-facts> / <vault-routing>). Budget/results/timeout/threshold driven by CLAWMEM_PROFILE.
postcompact-inject SessionStart (compact) re-injects authoritative state after compaction → <vault-postcompact>
curator-nudge SessionStart surfaces curator actions; nudges when the report is stale
precompact-extract PreCompact extracts decisions / file paths / open questions before compaction
decision-extractor Stop LLM → observations + causal links + contradiction detection + SPO triples
handoff-generator Stop LLM session summary → handoffs
feedback-loop Stop tracks referenced notes → confidence boosts, co-activations, utility signals

Default behavior: read injected <vault-context> first; if sufficient, answer immediately.

Hook blind spots (by design): hooks filter _clawmem/ artifacts, enforce score thresholds, and cap token budget — absence in <vault-context> does NOT mean absence in memory. If expected memory wasn't surfaced, escalate to Tier 3. Note the MCP retrieval tools themselves exclude _clawmem by default since v0.21.0 — pass includeInternal: true when system-internal memory (observations/handoffs/deductions) is the target.

Profiles: speed / balanced (default) / deep set the kept-score ratio (65% / 55% / 45%) and an activation floor. Only deep adds query expansion + reranking to the hook path. Profile and the hook timeout are set in ~/.claude/settings.json — see Operational gotchas for timeout tuning.

Tier 3 — Agent-initiated retrieval (MCP tools)

3-rule escalation gate

Escalate to MCP tools ONLY when one of these fires:

  1. Low-specificity injection — <vault-context> is empty or lacks the specific fact the task requires.
  2. Cross-session question — "why did we decide X", "what changed since last time", "when did we start Y".
  3. Pre-irreversible check — about to make a destructive / hard-to-reverse change; check the vault for prior decisions first.

All other retrieval is handled by Tier 2 hooks. Do NOT call MCP tools speculatively.

Tool routing

PREFERRED: memory_retrieve(query) — auto-classifies and routes to the optimal backend (query / intent_search / session_log / find_similar / query_plan). Use this instead of manually choosing.

1a. General recall      -> query(query, compact=true, limit=20)
    Full hybrid: BM25 + vector + expansion + deep rerank. Supports compact, collection,
    intent, candidateLimit. BM25 strong-signal bypass skips expansion when top hit >= 0.85
    with gap >= 0.15 (disabled when intent is provided).
1b. Causal/why/when/entity -> intent_search(query, enable_graph_traversal=true)
    MAGMA intent classification + intent-weighted RRF + multi-hop graph traversal.
    Use DIRECTLY (not as a fallback) for "why" / "when" / "how did X lead to Y" / entity links.
    Override: force_intent="WHY"|"WHEN"|"ENTITY"|"WHAT".
    (1a vs 1b are parallel options, chosen by query type — not sequential.)
1c. Multi-topic         -> query_plan(query, compact=true)
    Decomposes into 2-4 typed clauses (bm25/vector/graph), runs them in parallel, merges via RRF.
2.  Progressive disclosure -> multi_get("path1,path2") for full content of top hits
3.  Spot checks         -> search(query) (BM25, 0 GPU)  or  vsearch(query) (vector, 1 GPU)
4.  Chain tracing       -> find_causal_links(docid, direction="both", depth=5)
5.  Entity facts        -> kg_query(entity)  (SPO triples; different from intent_search's reasoning chains)
6.  Temporal context    -> timeline(docid, before=5, after=5)

All MCP tools

Tool Purpose
memory_retrieve Preferred. Auto-classifies + routes. Use instead of choosing manually.
query Full hybrid (BM25 + vector + rerank). General-purpose. WRONG for "why" (→ intent_search) or cross-session (→ session_log).
intent_search "why did we decide X" / "what caused Y" / "who worked on Z". Classifies intent, traverses graph edges — returns decision chains query can't find.
query_plan Multi-topic queries ("X and also Y", "compare A with B"). Splits + routes each clause.
search BM25 keyword — exact terms, config names, error codes. Fast, 0 GPU.
vsearch Vector semantic — conceptual/fuzzy when vocabulary unknown. ~100ms, 1 GPU.
get / multi_get Single doc by path/#docid / multiple by glob or comma-list.
find_similar "what else relates to X" — k-NN vector neighbors beyond keyword overlap.
find_causal_links Trace decision chains ("what led to X") over observation docs.
kg_query Entity SPO triples with temporal validity. Entity facts, NOT causal "why" (use intent_search).
session_log "last time" / "yesterday" / "what did we do". Do NOT use query for cross-session.
profile User profile (static facts + dynamic context).
memory_pin Lifecycle retention + priority among relevance-equivalent results (+0.3 composite boost on composite surfaces; exact-tie precedence on raw routes — vector + search non-recency). Use PROACTIVELY for constraints, architecture decisions, corrections.
memory_snooze Use PROACTIVELY when <vault-context> surfaces noise — snooze 30 days.
memory_forget Deactivate a memory by closest match. Sparingly — prefer snooze. Weak matches return a disambiguation list instead of acting (v0.23.0).
build_graphs Temporal backbone + semantic graph after bulk ingestion. NOT after every reindex. Reports N new edge(s), M total — 0 new on a rebuild is correct, not an empty graph.
timeline Temporal neighborhood around a doc. Progressive disclosure: search → timeline → get.
memory_evolution_status How a doc's A-MEM metadata evolved over time.
lifecycle_status / lifecycle_sweep / lifecycle_restore Lifecycle stats / archive stale (dry-run default) / restore auto-archived.
index_stats / status / reindex Doc counts + embedding coverage / quick health / force re-index (does NOT embed).
beads_sync / vault_sync / list_vaults Beads issues from Dolt / index a dir into a named vault / list vaults.

Multi-vault: all tools accept an optional vault param (omit for single-vault mode). Progressive disclosure: ALWAYS compact=true first → review snippets/scores → get / multi_get for full content.

Query optimization (4 levers)

The pipeline autonomously generates lex/vec/hyde variants, fuses BM25 + vector via RRF, and reranks with a cross-encoder — you do NOT choose search types. Your levers are tool selection, query string quality, intent, and candidateLimit.

Lever 1 — Tool selection (highest impact)

Pick the lightest tool that satisfies the need:

Tool Cost When
search(q, compact=true) BM25 only, 0 GPU Know exact terms, spot-check
vsearch(q, compact=true) Vector only, 1 GPU Conceptual/fuzzy, vocabulary unknown
query(q, compact=true) Full hybrid, 3+ GPU General recall, need best results
intent_search(q) Hybrid + graph Why/entity chains, when queries
query_plan(q, compact=true) Hybrid + decomposition Complex multi-topic

Lever 2 — Query string quality

The query string feeds BM25 (probes first, can short-circuit the pipeline) and anchors the 2×-weighted original signal in RRF — the single biggest determinant of result quality.

  • Keyword recall (BM25): 2–5 precise terms, no filler. Code identifiers work (handleError async). BM25 ANDs all terms as prefix matches (perf matches "performance") — no phrase search or negation. A strong hit (≥ 0.85, gap ≥ 0.15) skips expansion.
  • Semantic recall (vector): full natural-language question, be specific — "in the payment service, how are refunds processed" > "refunds".
  • Do NOT write hypothetical-answer-style queries — the expansion LLM already generates hyde variants; a long hypothetical dilutes BM25 and duplicates the pipeline.

Lever 3 — Intent (disambiguation)

Steers 5 autonomous stages (expansion, reranking, chunk selection, snippet extraction, strong-signal bypass). query("performance", intent="web page load times and Core Web Vitals").

  • Provide when: the term is polysemous in the vault, or the domain is known but the query alone is ambiguous.
  • Skip when: the query is already specific, single-domain vault, or using search/vsearch (intent only affects query).
  • Intent disables the BM25 strong-signal bypass (forces full expansion+rerank) — correct, since intent signals ambiguity.

Lever 4 — candidateLimit

How many RRF candidates reach the cross-encoder reranker (default 30). Lower (15) for high-confidence/speed/small-vault; higher (50) for broad topics/large vault/recall-over-speed.

Pipeline behavior

query (default Tier 3 workhorse)

Query + optional intent
  -> Temporal extraction (date ranges from "last week"/"March 2026")
  -> BM25 probe -> strong-signal check (skip expansion if top >= 0.85, gap >= 0.15; off when intent given)
  -> Query expansion (LLM text variants; intent steers the prompt)
  -> Parallel typed legs: BM25(orig) + Vector(orig) + BM25(lex exp) + Vector(vec/hyde exp) [+ temporal/entity if signalled]
  -> RRF (k=60; original lists get 2x positional weight, expanded 1x; top candidateLimit)
  -> Intent-aware chunk selection -> cross-encoder rerank (4000-char ctx; chunk dedup)
  -> rerank/RRF blend (0.9 reranker + 0.1 RRF tiebreaker; falls back to RRF if reranker down)
  -> composite scoring -> MMR diversity (Jaccard bigram > 0.6 demoted, not removed)

intent_search (specialist for causal chains)

Query -> intent classification (WHY/WHEN/ENTITY/WHAT)
  -> BM25 + Vector (intent-weighted RRF: BM25 for WHEN, vector for WHY)
  -> Graph traversal (WHY/ENTITY; multi-hop over memory_relations; outbound all edge types, inbound semantic+entity)
  -> cross-encoder rerank (200-char ctx) -> composite scoring

MPFP fusion is max-score, NOT RRF. The graph stage runs meta-path patterns ([semantic,causal], [entity,temporal], …) via Forward Push (α=0.15) and fuses by max-score ("best supporting path wins"), because propagation magnitude carries signal. This is distinct from the outer retrieval, which DOES fuse BM25+vector via RRF — two layers, two fusion rules, by design.

Key differences

Aspect query intent_search
Query expansion Yes (skipped on strong BM25) No
Intent intent param steers 5 stages Auto-detected (WHY/WHEN/ENTITY/WHAT)
Rerank context 4000 chars/doc 200 chars/doc
Graph traversal No Yes (WHY/ENTITY, multi-hop)
MMR diversity Yes No
compact / collection / candidateLimit Yes No
Best for most queries, progressive disclosure causal chains across docs

force_intent: WHY ("why", "what led to", "rationale", "tradeoff") · ENTITY (named component/person/service needing cross-doc linkage) · WHEN (timelines, first/last, "when did this change") — for WHEN start with enable_graph_traversal=false, fall back to query() if recall drifts.

Composite scoring (how ranking works)

Applied on the composite surfaces: hooks, query, and memory_retrieve's keyword/hybrid/causal/complex modes. v0.22.0: MCP vsearch and memory_retrieve semantic/discovery rank non-recency queries by RAW cosine instead (scoreBasis: "vector-cosine"; metadata breaks exact ties only; minScore filters raw with no default); recency-intent queries keep composite everywhere. v0.23.0: searchScore on FTS surfaces is the monotonic |bm25|/(1+|bm25|) transform (it was a constant 1.0 through v0.22.0 due to a clamp bug — keyword relevance contributed zero ordering); FTS-transform scores and cosines are independent monotonic signals, not one calibrated scale. v0.24.0: MCP search ranks non-recency queries by the RAW BM25 transform (scoreBasis: "fts-bm25"; metadata breaks exact ties only; minScore filters raw with no default) — judged keyword eval: raw MRR 0.848 vs composite 0.415 over 43 targets, composite losing even on the fresh-doc-favorable slice; recency-intent queries keep composite.

compositeScore = (0.50·searchScore + 0.25·recencyScore + 0.25·confidenceScore) × qualityMultiplier × coActivationBoost

Effective time (v0.27.0): recencyScore ages documents by authored_at ?? modified_at — mined/synthesized historical content ranks by when it was written, not when it was filed. Result metadata carries authored_at (null = unknown); temporal filters and recency-intent queries use the same axis.

  • qualityMultiplier = 0.7 + 0.6·qualityScore (0.7× penalty … 1.3× boost).
  • coActivationBoost = 1 + min(coCount/10, 0.15) (docs surfaced together get up to +15%).
  • Length normalization penalizes verbose entries (floor 30%); frequency boost capped at +10%.
  • Pinned docs: +0.3 additive on composite surfaces (capped at 1.0); on the raw routes (vector + search non-recency) pin = exact-tie precedence only.
  • query tool (v0.13.0+): non-recency queries use retrieval-tuned 0.70·search + 0.15·recency + 0.15·confidence. memory_retrieve's composite modes, context-surfacing, and search's recency branch keep the 0.50/0.25/0.25 default. (vsearch + memory_retrieve semantic/discovery use RAW cosine, and search uses the RAW BM25 transform, for non-recency queries — v0.22.0/v0.24.0: no composite weights at all.)
  • Recency intent ("latest"/"recent"/"last session") switches all to 0.10·search + 0.70·recency + 0.20·confidence.

Content-type half-lives: deductive / preference / hub / antipattern = ∞ (never decay) · decision 180d (very slow ranking decay — §36.11) · project 120d · research 90d · problem / milestone / note 60d · conversation / progress 45d · handoff 30d. Half-lives extend up to 3× for frequently-accessed memories. Attention decay: non-durable types (handoff, progress, conversation, note, project) lose 5% confidence/week without access; decision / deductive / preference / hub / research / antipattern are exempt.

→ full derivation: docs/concepts/composite-scoring.md.

Memory lifecycle (pin / snooze / forget — manual tools)

  • memory_pin (lifecycle retention + priority among relevance-equivalent results; +0.3 boost on composite surfaces, exact-tie precedence on raw routes) — PROACTIVELY when: user says "remember this"/"important"; an architecture/critical decision was just made; a user preference/constraint should persist across sessions. Do NOT pin routine/session-specific items.
  • memory_snooze — PROACTIVELY when a memory keeps surfacing but isn't relevant now, user says "not now"/"later", or content is time-boxed.
  • memory_forget — only when genuinely wrong or permanently obsolete. Prefer snooze for temporary suppression.
  • Contradiction auto-resolution: when decision-extractor detects a new decision contradicting an old one, the old one's confidence is lowered automatically (−0.25, floor 0.2) — no manual action needed. It stays retrievable; only its ranking drops. Removing it from retrieval outright (invalidated_at) is a separate, opt-in step behind CLAWMEM_CONTRADICTION_INVALIDATE, and applies only to content_type='observation' — a superseded decision is eroded, never retired, so do NOT tell a user that contradiction handling will retire a prior decision. Unarmed it logs WOULD invalidate and writes nothing. Do NOT suggest arming it without the vault-specific calibration in docs/guides/contradiction-invalidation.md.

Operational gotchas (agent-facing)

  • Empty context-surfacing → prompt < 20 chars (short memory-intent queries like "what did I say?" are exempt — they force retrieval), starts with /, or nothing scored above threshold. Check clawmem status (doc counts) + embedding coverage.
  • Vector search empty but BM25 works → missing embeddings (the watcher indexes but does NOT embed). Run clawmem embed or wait for the embed timer.
  • intent_search weak for WHY/ENTITY → sparse graph. Run build_graphs (temporal backbone + semantic edges). Otherwise don't run it after every reindex — A-MEM links per-doc automatically.
  • Rankings look RRF-flat / reranker suspect → clawmem rerank-health. A mis-served reranker (e.g. a GGUF that drops the score head) returns HTTP 200 but inert, non-discriminating scores, silently collapsing ranking to RRF. The reranker is a served sidecar, not a bundled model — verify it discriminates, don't assume liveness = correctness.
  • Intermittent UserPromptSubmit hook timed out after 8s — output discarded → almost always the context-surfacing hook's cold-start, NOT inference: a fresh Bun process + opening a large index.sqlite + a cold OS page cache. Warm calls are sub-second. On a memory-constrained host (e.g. WSL with a low memory cap) or a large vault, the cache is evicted between turns so it recurs on certain turns. A timed-out hook silently drops that turn's <vault-context> (degraded recall, no error). Durable fix: give the host enough RAM to keep the index + Bun modules cached; raising the hook timeout in ~/.claude/settings.json (8s default; no CLI knob) is only a secondary margin — avoid 15s+ as a standing default since the hook blocks prompt submission. Full detail: docs/troubleshooting.md → Hooks slow or near timeout / Tuning the context-surfacing hook timeout.
  • A known document is absent from search/vsearch/query but get by path returns it → it is invalidated (documents.invalidated_at IS NULL is a hard predicate on the FTS and both vector joins, with no query-time signal). On documents the only writer is contradiction invalidation, and only when armed — the invalidated_at in consolidation.ts is a different table. Diagnose + restore: docs/troubleshooting.md.
  • Anything setup-shaped (download blocked, server unreachable, watcher memory bloat, indexer bugs) → docs/troubleshooting.md. This skill does not duplicate it.

Anti-patterns

  • ❌ Manually pick query/intent_search/search when memory_retrieve can auto-route → ✅ memory_retrieve first.
  • ❌ Call MCP tools every turn → ✅ only when the 3-rule gate fires.
  • ❌ Re-search what's already in <vault-context>.
  • ❌ Run status routinely → ✅ only when retrieval feels broken or after large ingestion.
  • ❌ Pin everything → ✅ pin only persistent high-priority items.
  • ❌ Forget memories to "clean up" → ✅ let decay + contradiction detection handle it.
  • ❌ build_graphs after every reindex → ✅ only after bulk ingestion or when graph traversal is weak.
  • ❌ diary_write in Claude Code → ✅ hooks capture this automatically (diary is for non-hooked envs only).
  • ❌ kg_query for causal "why" → ✅ intent_search (kg_query is entity facts, not reasoning chains).

Curator agent

Maintenance agent for Tier-3 work the main agent neglects. Invoke: "curate memory" / "run curator" / "memory maintenance". Six phases: (1) health snapshot, (2) lifecycle triage (pin/snooze/propose-forget — never auto-confirms), (3) retrieval health probes, (4) reflect + consolidate --dry-run, (5) conditional graph rebuild, (6) collection hygiene. Safety rails: never auto-confirms forget, never runs embed, never edits config.

Tool selection (one-liner)

memory_retrieve(query) | query(compact=true) | intent_search(why/when/entity) | query_plan(multi-topic) -> multi_get -> search/vsearch (spot checks)

Setup / config / internals → AGENTS.md + docs/

This skill is operations-only. For installation, inference-server setup (the embedding/LLM/reranker services — the SOTA reranker is a seq-cls sidecar, not a GGUF), environment variables, systemd units, indexing/collection config, graph internals, and the OpenClaw (kind: memory) / Hermes (MemoryProvider) plugins, see AGENTS.md and docs/:

本页内容由上游仓库自动镜像，如有疑问请访问原始仓库。