config-reference.md

Config Reference

All settings live in openclaw.json under plugins.entries.openclaw-engram.config.

Use openclaw engram config-review for opinionated tuning recommendations and openclaw engram doctor for runtime or configuration problems. The narrative sections below explain the major feature groups; the schema-complete appendix at the bottom is the authoritative default-and-recommended matrix for every shipped config key.

Core

Setting	Default	Description
openaiApiKey	(env fallback in plugin mode)	Optional OpenAI API key, ${ENV_VAR} reference, or false to disable direct OpenAI entirely. When modelSource is gateway, Remnic does not inherit OPENAI_API_KEY; gateway provider auth is used instead.
openaiBaseUrl	(env fallback)	Override OpenAI API base URL (e.g. for proxies or compatible endpoints); falls back to OPENAI_BASE_URL env var
model	gpt-5.5	OpenAI model for extraction and consolidation
reasoningEffort	low	none, low, medium, high
memoryDir	~/.openclaw/workspace/memory/local	Memory storage root
workspaceDir	~/.openclaw/workspace	Workspace root (IDENTITY.md location)
captureMode	implicit	Memory write policy: implicit, explicit, or hybrid
debug	false	Enable debug logging

OpenClaw installs default new Remnic entries to modelSource: "gateway" so LLM calls use the gateway agent model chain instead of requiring a Remnic-specific OpenAI API key.

captureMode behavior:

• implicit: normal extraction/write behavior.
• explicit: normal conversation turns never create memories; only structured explicit capture writes or queues review items.
• hybrid: explicit capture writes immediately, while the normal extraction pipeline remains available.

Memory OS Presets

Setting	Default	Description
memoryOsPreset	(unset)	Optional advanced preset: conservative, balanced, research-max, or local-llm-heavy. Preset values seed the advanced config surface before explicit per-setting overrides are applied.

Preset intent:

• conservative keeps recall budgets lower and leaves experimental learning/graph features off.
• balanced enables the recommended indexing, artifact, and rerank defaults without turning on the higher-churn learning loops.
• research-max enables the broadest shipped experimental surface, including graph recall and adaptive policy loops.
• local-llm-heavy biases extraction/rerank/tooling toward local OpenAI-compatible endpoints and the fast local tier.

Backward compatibility note:

• memoryOsPreset: "research" is accepted as an alias for research-max, but new configs should use research-max.

Access Layer

Setting	Default	Description
agentAccessHttp.enabled	false	Start a local authenticated Remnic HTTP API during plugin startup
agentAccessHttp.host	127.0.0.1	Loopback bind host for the Remnic HTTP API
agentAccessHttp.port	4318	Bind port for the Remnic HTTP API (0 = ephemeral port)
agentAccessHttp.authToken	OPENCLAW_REMNIC_ACCESS_TOKEN / OPENCLAW_ENGRAM_ACCESS_TOKEN	Bearer token for the local HTTP API. Accepts a literal string (with ${ENV_VAR} expansion) or — under OpenClaw — a SecretRef object such as {"source":"exec","provider":"kc_openclaw_remnic_token","id":"value"} resolved at startup via the gateway secret resolver (issue #757). Standalone Remnic accepts strings only.
agentAccessHttp.maxBodyBytes	131072	Maximum accepted JSON request body size

When agentAccessHttp.enabled is on (or openclaw engram access http-serve is running), the same loopback server also serves the browser-based admin console shell at /engram/ui/. The shell is static, ships with packaged plugin builds, and still requires the configured bearer token over /engram/v1/... for memory data and operator actions.

Access-layer safety notes:

• HTTP startup fails closed when no bearer token is configured.
• Request bodies are capped by agentAccessHttp.maxBodyBytes.
• Explicit write routes are rate-limited and support schemaVersion, idempotencyKey, and dryRun envelopes.
• The stdio MCP server (openclaw engram access mcp-serve) uses the same internal access service as HTTP, so recall/read/write behavior stays aligned across both transports.
• MCP is intentionally zero-config on the Engram side: launch openclaw engram access mcp-serve from the client and it will use the same local memory directory, namespace rules, and explicit-capture policy as the in-process plugin runtime.

Buffer & Triggers

Setting	Default	Description
triggerMode	smart	smart, every_n, or time_based
bufferMaxTurns	5	Max buffered turns before forced extraction
bufferMaxMinutes	15	Max minutes before forced extraction
highSignalPatterns	[]	Additional regex patterns for immediate extraction
consolidateEveryN	3	Run consolidation every N extractions

Extraction Guardrails

Setting	Default	Description
extractionDedupeEnabled	true	Skip extraction if the same buffer was already extracted recently
extractionDedupeWindowMs	300000	Dedup window in milliseconds (default 5 minutes)
extractionMinChars	40	Minimum buffer character count to trigger extraction
extractionMinUserTurns	1	Minimum user turns in buffer before extraction
extractionMaxTurnChars	4000	Truncate each turn to this many chars before sending to LLM
extractionMaxFactsPerRun	12	Cap on facts extracted per LLM call
extractionMaxEntitiesPerRun	6	Cap on entities extracted per LLM call
extractionMaxQuestionsPerRun	3	Cap on curiosity questions generated per LLM call
extractionMaxProfileUpdatesPerRun	4	Cap on profile update statements per LLM call
beforeResetTimeoutMs	2000	Max time (ms, clamped to [100, 30000]) to wait for a reset-triggered flush before returning control to the host. Operators running a local LLM for extraction often want this higher — a 7B model on CPU can take 2–5s per extraction, and the default can abort the queued follow-up flush before it completes. See issue #549 for the error-vs-debug log-level behavior around these aborts.

Search Backend (v9.0)

Setting	Default	Description
searchBackend	"qmd"	Search engine to use: "qmd", "orama", "lancedb", "meilisearch", "remote", "noop"
lanceDbPath	{memoryDir}/lancedb	LanceDB database directory
lanceEmbeddingDimension	1536	Vector dimension for LanceDB
meilisearchHost	http://localhost:7700	Meilisearch server URL
meilisearchApiKey	(none)	Meilisearch API key
meilisearchTimeoutMs	30000	Meilisearch request timeout
meilisearchAutoIndex	false	Auto-push documents to Meilisearch on update
oramaDbPath	{memoryDir}/orama	Orama database directory
oramaEmbeddingDimension	1536	Vector dimension for Orama
remoteSearchBaseUrl	http://localhost:8181	Remote search service URL
remoteSearchApiKey	(none)	Remote search API key
remoteSearchTimeoutMs	30000	Remote search request timeout

See Search Backends for detailed configuration and comparison.

Retrieval & Recall Budget

Setting	Default	Description
recallBudgetChars	maxMemoryTokens * 4	Total character budget for assembled recall context. Controls how much memory context is injected into agent prompts. If unset, falls back to maxMemoryTokens * 4. See Recall Budget Tuning below.
maxMemoryTokens	2000	Legacy token cap. Only used to compute recallBudgetChars when that setting is absent. Prefer setting recallBudgetChars directly.
qmdEnabled	true	Use QMD for hybrid search
qmdCollection	openclaw-engram	QMD collection name
qmdMaxResults	8	Final result cap after over-scanning and ranking (fetch size may be larger)
qmdColdTierEnabled	false	Query a secondary cold QMD collection before archive fallback when hot recall misses
qmdColdCollection	openclaw-engram-cold	QMD collection name used for cold-tier recall
qmdColdMaxResults	8	Final result cap for cold-tier recall before merging into the normal ranking pipeline
qmdPath	(auto)	Absolute path to qmd binary (bypasses PATH)
qmdSupportedVersion	2.5.3	Highest QMD version this Remnic build will auto-install
qmdAutoUpgradeEnabled	false	Opt-in auto-upgrade for PATH/fallback QMD installs; explicit qmdPath is never overwritten
qmdAutoUpgradeCheckIntervalMs	86400000	Minimum interval between auto-upgrade attempts
qmdChunkStrategy	auto	QMD chunk strategy to forward when the installed QMD supports it (auto or regex)
qmdCandidateLimit	(none)	Optional QMD candidate limit forwarded to supported QMD query paths
qmdQueryRerankEnabled	true	Set false to ask QMD to skip its built-in rerank step when supported
qmdIndexName	(none)	Optional QMD named index forwarded as qmd --index <name> ... when QMD 2.5+ supports named index selection. Leave unset during upgrades unless existing QMD data is already in that named index; QMD's default data lives in ~/.cache/qmd/index.sqlite, and changing this can point Remnic at an empty DB.
qmdForceCpu	false	Set QMD_FORCE_CPU=1 for QMD child processes to bypass GPU probing
qmdGpuBackend	(none)	Optional QMD_LLAMA_GPU override (auto, metal, cuda, vulkan, or false)
qmdEmbedParallelism	(none)	Optional QMD_EMBED_PARALLELISM override, clamped to 1-8
qmdEmbedModel	(none)	Optional QMD_EMBED_MODEL override used by QMD indexing and vector search
qmdRerankModel	(none)	Optional QMD_RERANK_MODEL override used by QMD reranking
qmdGenerateModel	(none)	Optional QMD_GENERATE_MODEL override used by QMD query expansion
qmdDaemonEnabled	true	Prefer QMD MCP daemon for recall/search when available (lower contention); fail-open to subprocess search/hybrid paths
qmdDaemonUrl	http://localhost:8181/mcp	Legacy compatibility setting; current runtime uses shared stdio qmd mcp rather than the HTTP endpoint directly
qmdDaemonRecheckIntervalMs	60000	Interval to re-probe daemon availability after failure
qmdIntentHintsEnabled	false	Forward inferred recall intent into QMD unified search when supported
qmdExplainEnabled	false	Capture QMD explain traces in state/last_qmd_recall.json and memory_qmd_debug
embeddingFallbackEnabled	true	Use embedding search when QMD is unavailable
embeddingFallbackProvider	auto	auto, openai, or local — selects embedding API for fallback
recordEmptyRecallImpressions	false	If true, write recall impression rows with empty memoryIds when no memory context is injected
knowledgeIndexEnabled	true	Inject entity/topic index into recall context
knowledgeIndexMaxEntities	40	Max entities included in the knowledge index
knowledgeIndexMaxChars	4000	Max characters of knowledge index injected
entityRetrievalEnabled	true	Enable entity-oriented recall hints for who is, what do we know about, and transcript-backed recent-turn pronoun follow-ups within the active recall namespace
entityRetrievalMaxChars	2400	Max characters injected by the entity retrieval section
entityRetrievalMaxHints	2	Max entity targets summarized in a single recall pass
entityRetrievalMaxSupportingFacts	6	Max direct-answer supporting facts/timeline snippets considered per target
entityRetrievalMaxRelatedEntities	3	Max related entities listed per target when confidence is high
entityRetrievalRecentTurns	6	Number of recent transcript turns scanned for pronoun carry-forward and short follow-up resolution
entityRelationshipsEnabled	true	Persist entity-relationship edges that power direct-answer recall summaries
entityActivityLogEnabled	true	Keep per-entity recent-activity snippets for answer synthesis
entityActivityLogMaxEntries	20	Max recent activity entries retained per entity
entityAliasesEnabled	true	Track normalized aliases for entity resolution and merge safety
entitySummaryEnabled	true	Maintain synthesized entity summaries used by retrieval and tooling
recallBudgetChars	maxMemoryTokens * 4	Hard cap for total assembled recall context (final safety trim before system prompt injection)
recallPipeline	(built-in ordered defaults)	Ordered section controls for recall assembly, including per-section caps and knobs
recallDirectAnswerEnabled	false	Opt in to the direct-answer retrieval tier (issue #518). When enabled, the tier runs in observation mode: it annotates recall tier explain data without short-circuiting the QMD path. See Retrieval Explain.
recallDirectAnswerTokenOverlapFloor	0.55	Minimum query↔memory token-overlap ratio required for direct-answer eligibility. Set to 0 to disable the gate.
recallDirectAnswerImportanceFloor	0.7	Minimum calibrated importance score required for direct-answer eligibility. Set to 0 to disable the gate. verificationState: "user_confirmed" bypasses this check.
recallDirectAnswerAmbiguityMargin	0.15	If the second-best candidate scores within this ratio of the top, direct-answer defers to the hybrid tier.
recallDirectAnswerEligibleTaxonomyBuckets	["decisions","principles","conventions","runbooks","entities"]	Taxonomy category IDs eligible for direct-answer routing. Set to [] to disable the gate without unsetting enabled.

recallPipeline entries

recallPipeline is an array of section entries:

{
  "id": "knowledge-index",
  "enabled": true,
  "maxChars": 3000,
  "maxEntities": 25
}

Supported keys:

Key	Type	Notes
id	string	Section identifier (required)
enabled	boolean	Enable/disable the section
maxChars	number | null	Per-section char cap (null = uncapped by section)
maxHints	number	entity-retrieval section only; max resolved entity targets
maxSupportingFacts	number	entity-retrieval section only; direct-answer evidence budget per target
maxRelatedEntities	number	entity-retrieval section only; related-entity cap per target
consolidateTriggerLines	number	profile section only; profile consolidation trigger line count
consolidateTargetLines	number	profile section only; consolidation target line count
maxEntities	number	knowledge-index section only; per-section entity cap
maxResults	number	memories section only; cap injected memory result count
recentTurns	number	entity-retrieval section only; transcript follow-up window
maxTurns	number	transcript section only
maxTokens	number	transcript section only
lookbackHours	number	transcript / summaries section only
maxCount	number	summaries section only
topK	number	conversation-recall section only
timeoutMs	number	conversation-recall section only
maxPatterns	number	compounding section only

Recall Budget Tuning

The recall budget controls how much context Engram injects into each agent prompt. Getting this right is critical — too small and memories are silently truncated; too large and you waste context window space.

How it works (v9.0.66+): Engram assembles recall context in pipeline section order (shared-context → profile → entity retrieval → knowledge index → ... → memories → transcripts → summaries). The budget-aware assembler reserves space for the memories section so earlier sections cannot fully exhaust the budget. However, the reservation is minimal (heading-sized). If the total budget is too small, earlier sections still crowd out memory content.

Common pitfall: The default budget is maxMemoryTokens * 4 = 8,000 chars. A typical profile is 4,000–8,000 chars and shared context adds another 4,000–6,000 chars. With these defaults, the memories section is still included (it is a protected section), but may be truncated to heading-only (~24 chars) with no actual memory content. The lastRecall state file will show successful memory retrieval (non-empty memoryIds) but the agent sees only the section heading because the content was truncated during context assembly.

Recommended values:

Model context window	Suggested recallBudgetChars	Reasoning
8K–16K tokens	16000	Tight budget; consider capping profile via recallPipeline
32K–128K tokens	32000–64000	Room for all sections including memories
200K+ tokens (Claude Opus/Sonnet, GPT-5)	64000–128000	Generous budget; 16K–32K tokens is a small fraction of context

Example config for large-context models:

{
  "recallBudgetChars": 64000
}

Diagnosing budget exhaustion: Check ~/.openclaw/workspace/memory/local/state/last_recall.json. Each session entry records includedSections, finalContextChars, and memoryIds. Because memories is a protected section, it is always included — but under tight budgets it may be truncated to heading-only. If memoryIds is non-empty but finalContextChars is close to the budget and the memories section content is missing or minimal, the budget was too small and memories were retrieved but truncated during assembly.

Capping individual sections: You can override the recallPipeline to add maxChars to any section:

{
  "recallPipeline": [
    { "id": "shared-context", "enabled": true, "maxChars": 4000 },
    { "id": "profile", "enabled": true, "maxChars": 4000 },
    { "id": "entity-retrieval", "enabled": true, "maxChars": 2400 },
    { "id": "knowledge-index", "enabled": true, "maxChars": 4000 },
    { "id": "memories", "enabled": true }
  ]
}

Note: recallPipeline controls ordering and can explicitly disable sections via "enabled": false. Unlisted sections default to enabled and are appended after the listed entries. To exclude a section, include it with "enabled": false rather than omitting it.

Coding Mode

Setting	Default	Description
codingMode.projectScope	true	Auto-scope memory to the git project (stable origin-URL hash). Set to false to disable project-based namespace isolation.
codingMode.branchScope	false	Additionally overlay the current branch on top of the project namespace. Project-level reads remain visible through readFallbacks.
codingMode.globalFallback	true	Include the root/global namespace in recall read-fallbacks for project-scoped sessions. Global facts (framework bugs, library behavior, user preferences) surface across all projects. Set to false for strict project isolation.
extractionScopeClassificationEnabled	true	Classify extracted facts as "global" or "project" scope. Global facts are promoted to the shared root namespace so they are visible across all projects.

See Coding agent mode for full details on project detection, cwd auto-resolution, projectTag for non-git sessions, and cross-project knowledge sharing.

Native Knowledge

Setting	Default	Description
nativeKnowledge.enabled	false	Enable curated-file and adapter-backed native knowledge recall.
nativeKnowledge.includeFiles	["IDENTITY.md","MEMORY.md"]	Workspace-relative markdown files to chunk into the native knowledge recall section and track incrementally in backend-agnostic sync state.
nativeKnowledge.maxChunkChars	900	Maximum chunk size before heading/paragraph-aware splitting.
nativeKnowledge.maxResults	4	Maximum native knowledge chunks injected into recall.
nativeKnowledge.maxChars	2400	Maximum total characters injected by the native knowledge section.
nativeKnowledge.stateDir	state/native-knowledge	memoryDir-relative directory used for backend-agnostic adapter sync state.
nativeKnowledge.openclawWorkspace	unset	Optional OpenClaw workspace adapter for bootstrap docs, handoffs, daily summaries, and automation notes.
nativeKnowledge.obsidianVaults	[]	Optional Obsidian vault adapters to sync into native knowledge recall.

nativeKnowledge.openclawWorkspace

Key	Default	Description
enabled	false	Enable the OpenClaw workspace artifact adapter.
bootstrapFiles	["IDENTITY.md","MEMORY.md","USER.md"]	Workspace-relative bootstrap docs treated as high-confidence native knowledge.
handoffGlobs	["**/*handoff*.md","handoffs/**/*.md"]	Workspace-relative globs used to discover handoff notes.
dailySummaryGlobs	["**/*daily*summary*.md","summaries/**/*.md"]	Workspace-relative globs used to discover daily summary notes.
automationNoteGlobs	[]	Optional workspace-relative globs for automation-written status or operating notes.
workspaceDocGlobs	[]	Optional workspace-relative globs for other explicitly allowlisted workspace docs.
excludeGlobs	[]	Additional excludes appended to the built-in safety exclusions (.git/**, node_modules/**, dist/**, build/**, coverage/**, **/*.log, **/.env*, **/*.pem, **/*.key).
sharedSafeGlobs	[]	Optional workspace-relative globs tagged as shared_safe when no explicit privacy class is present.

nativeKnowledge.obsidianVaults entries

Each vault entry supports:

Key	Default	Description
id	vault-{n}	Stable adapter identifier used in synced metadata and recall formatting.
rootDir	required	Absolute path to the Obsidian vault root.
includeGlobs	["**/*.md"]	Vault-relative globs eligible for sync.
excludeGlobs	[".obsidian/**","**/*.canvas","**/*.png","**/*.jpg","**/*.jpeg","**/*.gif","**/*.pdf"]	Vault-relative globs excluded from sync.
namespace	unset	Default namespace assigned to synced notes from this vault.
privacyClass	unset	Operator-defined privacy classification preserved on synced note chunks.
folderRules	[]	Optional per-folder overrides for namespace and privacy class. Longest matching prefix wins.
dailyNotePatterns	["YYYY-MM-DD"]	Filename patterns used to derive a note date from the vault-relative path.
materializeBacklinks	false	When enabled, compute backlinks from wikilink targets and expose them in recall metadata.

Example:

{
  "nativeKnowledge": {
    "enabled": true,
    "includeFiles": ["IDENTITY.md", "MEMORY.md", "TEAM.md"],
    "openclawWorkspace": {
      "enabled": true,
      "bootstrapFiles": ["IDENTITY.md", "MEMORY.md", "USER.md"],
      "handoffGlobs": ["handoffs/**/*.md"],
      "dailySummaryGlobs": ["summaries/**/*.md"],
      "automationNoteGlobs": ["automation/**/*.md"],
      "sharedSafeGlobs": ["automation/shared/**/*.md"]
    },
    "obsidianVaults": [
      {
        "id": "personal",
        "rootDir": "/Users/you/Documents/Obsidian",
        "namespace": "shared",
        "privacyClass": "private",
        "folderRules": [
          { "pathPrefix": "Projects", "namespace": "work", "privacyClass": "team" }
        ],
        "dailyNotePatterns": ["Daily/YYYY-MM-DD", "YYYY-MM-DD"],
        "materializeBacklinks": true
      }
    ]
  }
}

Direct includeFiles sync plus the OpenClaw workspace adapter both persist incremental sync state and tombstones under nativeKnowledge.stateDir, preserve source metadata on each chunk when derivable, and dedupe exact overlaps so enabling the adapter does not double-inject bootstrap docs.

v8.0 Memory OS

Setting	Default	Description
recallPlannerEnabled	true	Lightweight retrieve-vs-think gating
recallPlannerLlmEnabled	false	Opt in to LLM-based recall planning (issue #1367). When off, recall mode is decided by the regex heuristic. When on, recallPlannerModel classifies recall intent via the gateway/fallback LLM chain (provider-agnostic) and falls back to the heuristic on timeout/error. Requires recallPlannerEnabled. See Heuristic vs LLM planning.
recallPlannerMaxQmdResultsMinimal	4	QMD cap in minimal recall mode
memoryBoxesEnabled	false	Enable Memory Box topic-windowed grouping
traceWeaverEnabled	false	Link recurring-topic boxes into named traces
boxTimeGapMs	1800000	Milliseconds of inactivity that seal an open box (default 30 min)
boxTopicShiftThreshold	0.35	Topic overlap below this seals the box
boxMaxMemories	50	Max memories before forced seal
traceWeaverLookbackDays	7	Days to look back for matching traces
traceWeaverOverlapThreshold	0.4	Minimum topic overlap to join an existing trace
boxRecallDays	3	Days of boxes to inject into recall context
episodeNoteModeEnabled	false	Classify memories as episode or note
verbatimArtifactsEnabled	false	Store high-confidence memories as verbatim anchors
verbatimArtifactsMinConfidence	0.8	Minimum confidence for artifact writes
verbatimArtifactsMaxRecall	5	Max artifact anchors injected per recall
verbatimArtifactCategories	["decision","correction","principle","commitment"]	Eligible categories
intentRoutingEnabled	false	Write intent metadata; boost compatible recalls
intentRoutingBoost	0.12	Max additive score boost from intent compatibility

v8.1 Temporal + Tag Indexes

Setting	Default	Description
queryAwareIndexingEnabled	false	Build and maintain temporal (state/index_time.json) and tag (state/index_tags.json) indexes after each extraction. Enables score boosts for temporal queries and #tag tokens at recall time.
queryAwareIndexingMaxCandidates	200	Max candidate paths from the index prefilter (0 = no cap).

v8.3 Lifecycle Policy Engine

Setting	Default	Description
lifecyclePolicyEnabled	false	Enable lifecycle scoring + transitions + retrieval weighting.
lifecycleFilterStaleEnabled	false	Filter lifecycle stale/archived candidates from retrieval before final cap (only when policy is enabled).
lifecyclePromoteHeatThreshold	0.55	Heat threshold for promotion toward validated/active.
lifecycleStaleDecayThreshold	0.65	Decay threshold to move a memory to stale.
lifecycleArchiveDecayThreshold	0.85	Decay threshold to move a memory to archived (non-protected categories).
lifecycleProtectedCategories	["decision","principle","commitment","preference","procedure"]	Categories protected from automatic archive transition (includes procedure when procedural memories exist).
lifecycleMetricsEnabled	false (auto-true when policy enabled unless explicitly set)	Emit lifecycle metrics snapshot at state/lifecycle-metrics.json.

v8.3 Proactive + Policy Learning Foundation

Setting	Default	Description
proactiveExtractionEnabled	false	Enable proactive extraction second-pass paths (feature-gated).
contextCompressionActionsEnabled	false	Enable context compression action tool paths and action telemetry wiring.
compressionGuidelineLearningEnabled	false	Enable adaptive compression guideline learning loop.
maxProactiveQuestionsPerExtraction	2	Hard cap on proactive self-questions per extraction (0 disables).
proactiveExtractionTimeoutMs	2500	Hard timeout for proactive question generation plus bounded answer synthesis (0 disables the second pass).
proactiveExtractionMaxTokens	900	Token budget applied to each proactive extraction sub-call (0 disables the second pass).
proactiveExtractionCategoryAllowlist	unset	Optional category allowlist for proactive second-pass writes; when set, lower-confidence or off-category proactive facts are dropped before persistence.
maxCompressionTokensPerHour	1500	Hourly token budget for compression-learning workflows (0 disables).

v8.3 Tool + State Artifacts

• context_checkpoint tool:
  • gated by contextCompressionActionsEnabled
  • records append-only telemetry in state/memory-actions.jsonl
• memory_action_apply tool:
  • gated by contextCompressionActionsEnabled
  • records append-only action + outcome telemetry in state/memory-actions.jsonl
• compressionGuidelineLearningEnabled:
  • consolidation synthesizes/updates state/compression-guidelines.md
  • optimizer metadata/version state persists to state/compression-guideline-state.json
  • synthesis is fail-open and never blocks consolidation
• proactiveExtractionTimeoutMs / proactiveExtractionMaxTokens:
  • bound both proactive self-question generation and the same-buffer answer-synthesis pass
  • 0 remains a hard disable for the proactive second pass
• proactiveExtractionCategoryAllowlist:
  • filters proactive second-pass facts before persistence so only allowlisted categories are emitted
  • does not affect the base extraction pass

v8.13 Action-Policy Rollout Presets

Use these as operator presets for progressive rollout. All are baseline-safe when disabled.

conservative:

{
  "contextCompressionActionsEnabled": false,
  "proactiveExtractionEnabled": false,
  "compressionGuidelineLearningEnabled": false,
  "compressionGuidelineSemanticRefinementEnabled": false,
  "proactiveExtractionTimeoutMs": 2500,
  "proactiveExtractionMaxTokens": 900,
  "maxCompressionTokensPerHour": 0
}

balanced:

{
  "contextCompressionActionsEnabled": true,
  "proactiveExtractionEnabled": true,
  "compressionGuidelineLearningEnabled": true,
  "compressionGuidelineSemanticRefinementEnabled": false,
  "proactiveExtractionTimeoutMs": 2500,
  "proactiveExtractionMaxTokens": 900,
  "maxCompressionTokensPerHour": 1500
}

research-max:

{
  "contextCompressionActionsEnabled": true,
  "proactiveExtractionEnabled": true,
  "compressionGuidelineLearningEnabled": true,
  "compressionGuidelineSemanticRefinementEnabled": true,
  "compressionGuidelineSemanticTimeoutMs": 2500,
  "proactiveExtractionTimeoutMs": 2500,
  "proactiveExtractionMaxTokens": 900,
  "maxCompressionTokensPerHour": 3000
}

Disabled-path compatibility guarantees:

• contextCompressionActionsEnabled=false keeps action tooling and action-policy telemetry inactive.
• proactiveExtractionTimeoutMs=0 or proactiveExtractionMaxTokens=0 keeps the proactive second pass fully disabled.
• maxCompressionTokensPerHour=0 remains a hard disable (no implicit non-zero coercion).
• compressionGuidelineLearningEnabled=false keeps consolidation behavior baseline-equivalent.

Budget Mapping Notes

The original v8 roadmap listed several operator knobs that are now split across the live config surface.

Roadmap knob	Live config surface
maxRecallTokens	maxMemoryTokens for token budget, plus recallBudgetChars for final assembled-context trimming.
maxRecallMs	No single global wall-clock cap. Use stage-specific limits such as recallPlannerTimeoutMs, conversationRecallTimeoutMs, and rerankTimeoutMs.
maxCompressionTokensPerHour	maxCompressionTokensPerHour
maxGraphTraversalSteps	maxGraphTraversalSteps
maxArtifactsPerSession	No dedicated per-session write cap. The nearest shipped controls are verbatimArtifactsEnabled, verbatimArtifactsMaxRecall, and verbatimArtifactCategories.
maxProactiveQuestionsPerExtraction	maxProactiveQuestionsPerExtraction
maxProactiveExtractionMs	proactiveExtractionTimeoutMs
maxProactiveExtractionTokens	proactiveExtractionMaxTokens
indexRefreshBudgetMs	Use refresh cadence + timeout controls such as qmdUpdateMinIntervalMs, qmdUpdateTimeoutMs, and conversationIndexMinUpdateIntervalMs.

v8.14 Hot/Cold Tier Parity + Migration

Setting	Default	Description
qmdTierMigrationEnabled	false	Enable value-aware migration between hot and cold QMD tiers.
qmdTierDemotionMinAgeDays	14	Minimum age (days) before a hot memory can be considered for demotion.
qmdTierDemotionValueThreshold	0.35	Value threshold at/below which hot memories are eligible for cold demotion.
qmdTierPromotionValueThreshold	0.7	Value threshold at/above which cold memories are eligible for hot promotion.
qmdTierParityGraphEnabled	true	Keep graph-assist behavior parity between hot and cold retrieval paths.
qmdTierParityHiMemEnabled	true	Keep HiMem episode/note handling parity between hot and cold retrieval paths.
qmdTierAutoBackfillEnabled	false	Enable automated cold-tier parity backfill jobs.

Gateway Model Source

Route all Engram LLM calls through the OpenClaw gateway's agent model chain instead of Engram's own openaiApiKey/localLlm* configuration. This lets you define a single fallback chain per agent persona in openclaw.json and reuse the gateway's provider credentials.

Setting	Default	Description
modelSource	gateway for new OpenClaw installs; plugin otherwise	gateway delegates to a gateway agent's model chain; plugin uses Engram's own openai/localLlm config
gatewayAgentId	""	Agent persona ID from openclaw.json → agents.list[] for primary LLM calls (extraction, consolidation, summarization). Falls back to agents.defaults.model if empty.
fastGatewayAgentId	""	Agent persona ID for fast-tier ops (rerank, entity summaries, compression guidelines). Uses gatewayAgentId chain when empty.
taskModelChain	(unset)	Optional inline { "primary", "fallbacks": [] } model chain for Remnic's background tasks (extraction, extraction judge, fact/profile/identity consolidation, summarization, semantic consolidation, calibration, causal consolidation). Resolves through gateway providers. When set, it overrides gatewayAgentId/agents.defaults.model for these tasks only. Requires modelSource: "gateway" — ignored (with a startup warning) in plugin mode.

When modelSource is gateway:

• localLlmEnabled and the direct OpenAI client are bypassed for primary LLM dispatch — all LLM calls flow through FallbackLlmClient with the configured agent chain
• Extraction and consolidation start on the configured gateway chain directly; the historical "falling back to gateway" wording only applies when Engram is still in plugin mode
• The existing openaiApiKey, model, and localLlm* settings are ignored for LLM dispatch but retained as config for backward compatibility; OPENAI_API_KEY is not inherited in gateway mode
• localLlmFast* settings are also bypassed when fastGatewayAgentId is set
• Reranking uses the fastGatewayAgentId chain (or gatewayAgentId if fast is unset) instead of the local LLM — this can dramatically reduce rerank latency when the fast chain points at a cloud provider

Task-specific model chain (taskModelChain)

By default, gateway-mode background work — extraction, the extraction judge, fact/profile/identity consolidation, summarization, semantic consolidation, calibration, and causal consolidation — shares the gatewayAgentId persona chain, or agents.defaults.model when no persona is set. That ties lightweight memory tasks to whatever chain the main agent uses, which is often larger/pricier than these tasks need. (The one exception: semanticConsolidationModel: "fast" keeps using fastGatewayAgentId — an explicit fast-tier choice is honored over taskModelChain.)

Set taskModelChain to give those tasks their own cheap/fast chain without defining a persona or touching agents.defaults.model:

{
  "modelSource": "gateway",
  "taskModelChain": {
    "primary": "zai/glm-4.7-flash",
    "fallbacks": ["fireworks/accounts/fireworks/models/glm-5p1"]
  }
}

Notes:

• Models resolve through the same models.providers auth/routing as everything else — only the chain differs.
• taskModelChain takes precedence over gatewayAgentId for these tasks; the main agent persona is unaffected.
• A reachable agents.defaults.model is appended as an implicit last resort to a taskModelChain only, so an exhausted task chain never blocks a flush. Persona/default chains are never augmented this way.
• It only applies in gateway mode. In plugin mode it is ignored and Remnic logs a startup warning. Plugin-mode users who hit non-OpenAI model-ID failures at the direct client should switch to modelSource: "gateway" and use taskModelChain (see issue #1365).
• A present-but-malformed value (missing primary, wrong types) is rejected at config-parse time rather than silently ignored.

Setup

1. Define providers in agents/main/agent/models.json with the endpoints and credentials you want Engram to use (e.g., fireworks, zai, anthropic, lmstudio).

2. Create agent personas in openclaw.json → agents.list[]:

{
  "id": "engram-llm",
  "default": false,
  "name": "Engram LLM Chain",
  "model": {
    "primary": "fireworks/accounts/fireworks/routers/kimi-k2p5-turbo",
    "fallbacks": [
      "zai/glm-5",
      "anthropic/claude-sonnet-4-6",
      "lmstudio/qwen3.5-35b-a3b-mlx-lm"
    ]
  }
},
{
  "id": "engram-llm-fast",
  "default": false,
  "name": "Engram Fast LLM Chain",
  "model": {
    "primary": "fireworks/accounts/fireworks/routers/kimi-k2p5-turbo",
    "fallbacks": [
      "zai/glm-5-turbo",
      "anthropic/claude-sonnet-4-6",
      "lmstudio/qwen3.5-35b-a3b-mlx-lm"
    ]
  }
}

Model strings use the format provider/model-id where provider matches a key in the providers object of your agent's models.json. Built-in OpenClaw providers (e.g., openai-codex, google-vertex, github-copilot) work automatically — they don't need explicit entries in models.json since the gateway materializes them from its plugin catalogs.

3. Configure Engram in openclaw.json → plugins.entries.openclaw-engram.config:

{
  "modelSource": "gateway",
  "gatewayAgentId": "engram-llm",
  "fastGatewayAgentId": "engram-llm-fast"
}

4. Restart the gateway for changes to take effect.

How the fallback chain works

When a primary model call fails (timeout, HTTP error, empty response), FallbackLlmClient tries each fallback in order. The chain stops at the first successful response.

Provider lookup checks the explicit models.providers config first, then falls back to the gateway's materialized models.json (~/.openclaw/agents/main/agent/models.json), which contains all providers including built-in ones registered by gateway plugins (e.g., openai-codex with OAuth, google-vertex, github-copilot). This means any provider the gateway knows about — including OAuth-based providers — can be used in Engram's model chain without additional configuration.

API key resolution

Provider auth is resolved using OpenClaw's native runtime. Engram first tries the gateway's getRuntimeAuthForModel() function, which handles all provider-specific transforms — OAuth token exchange (for openai-codex, github-copilot, etc.), base URL overrides, profile-based credentials, and secret reference formats — using the same codepath the gateway uses for its own agent sessions.

If the gateway runtime isn't available (e.g., running outside the gateway process), Engram falls back to resolveProviderApiKey() for secret ref resolution, then checks the PROVIDER_NAME_API_KEY environment variable before skipping the provider.

This means your existing auth setup works automatically — OAuth providers, API keys, 1Password, Vault, env vars, and plain-text keys all work without special Engram configuration.

Switching back

Set modelSource to plugin (or remove it) to restore the original behavior where Engram uses its own localLlm* and openaiApiKey settings.

Local LLM / OpenAI-Compatible Endpoint

Setting	Default	Description
localLlmEnabled	false	Enable Engram's local/compatible endpoint when modelSource remains plugin
localLlmUrl	http://localhost:1234/v1	Base URL for endpoint
localLlmModel	local-model	Model ID
localLlmApiKey	(unset)	Optional API key
localLlmHeaders	(unset)	Extra HTTP headers
localLlmAuthHeader	true	Send Authorization: Bearer header when key set
localLlmFallback	true	Fall back to gateway model chain on failure
localLlmTimeoutMs	180000	Total timeout for primary local extraction/consolidation calls
localLlmRetry5xxCount	1	Retry count for transient 5xx responses from the local endpoint
localLlmRetryBackoffMs	400	Base backoff in milliseconds for local endpoint retries
localLlm400TripThreshold	5	Consecutive 4xx responses before the local endpoint is temporarily tripped
localLlm400CooldownMs	120000	Cooldown window before retrying a tripped local endpoint
localLlmMaxContext	(unset)	Override context window size
localLlmFastEnabled	false	Enable a separate fast local tier for short planner/rerank/helper calls
localLlmFastModel	""	Optional model id for the fast local tier
localLlmFastUrl	http://localhost:1234/v1	Optional dedicated base URL for the fast local tier
localLlmFastTimeoutMs	15000	Timeout for the fast local tier
localLlmDisableThinking	true	Suppress chain-of-thought / thinking mode on the main local LLM by sending chat_template_kwargs: { enable_thinking: false } (issue #548). Structured-output tasks like extraction and consolidation gain nothing from reasoning tokens; thinking-capable models (Qwen 3.5, Gemma 4, DeepSeek) commonly blow the 60s timeout before emitting content. Set to false to restore thinking for narrative tasks. The fast-tier client always disables thinking regardless of this flag.
localLlmHomeDir	(unset)	Optional home-dir override used when resolving local helper binaries
localLmsCliPath	(auto)	Path to lms CLI (LM Studio)
localLmsBinDir	(auto)	LM Studio binary directory

v2 Features

Setting	Default	Description
identityEnabled	true	Enable agent identity reflections
injectQuestions	false	Inject open questions into system prompt
commitmentDecayDays	90	Days before fulfilled commitments are removed

v8.4 Identity Continuity

Setting	Default	Description
identityContinuityEnabled	false	Enable identity continuity workflows (anchor/incidents/audits)
identityInjectionMode	recovery_only	Identity context injection mode: recovery_only, minimal, full
identityMaxInjectChars	1200	Maximum identity continuity characters injected into recall
continuityIncidentLoggingEnabled	(follows identityContinuityEnabled when unset)	Explicit override for continuity incident logging
continuityAuditEnabled	false	Enable continuity audit generation workflows

v8.5 Active Session Observer

Setting	Default	Description
sessionObserverEnabled	false	Enable heartbeat observer checks for session growth-triggered extraction
sessionObserverDebounceMs	120000	Minimum milliseconds between observer-triggered extractions per session
sessionObserverBands	[{maxBytes:50000,triggerDeltaBytes:4800,triggerDeltaTokens:1200}, {maxBytes:200000,triggerDeltaBytes:9600,triggerDeltaTokens:2400}, {maxBytes:1000000000,triggerDeltaBytes:19200,triggerDeltaTokens:4800}]	Size-band thresholds used to trigger observer extraction when growth exceeds configured byte/token deltas

v8.5 Session Integrity + Recovery Ops

Session integrity diagnostics/repair are CLI-driven and intentionally config-light:

• openclaw engram session-check
• openclaw engram session-repair --dry-run|--apply

Safety contract:

• Repair defaults to dry-run.
• --apply only mutates Engram-managed transcript/checkpoint artifacts.
• OpenClaw session-file mutation requires explicit --allow-session-file-repair plus an explicit path and still does not perform automatic pointer rewiring.

v8.8 Live Graph Dashboard

Dashboard is an optional, separate process and not part of gateway hot-path config.

CLI defaults:

• openclaw engram dashboard start --host 127.0.0.1 --port 4319
• openclaw engram dashboard status
• openclaw engram dashboard stop

Operational safety:

• Bind to localhost by default.
• Explicitly choose non-loopback bind only when network controls are in place.

v8.7 Custom Memory Routing Rules

Setting	Default	Description
routingRulesEnabled	false	Enable write-time routing-rule evaluation for extracted facts
routingRulesStateFile	state/routing-rules.json	Relative state file path for persisted route rules

v2.2 Advanced Retrieval

See advanced-retrieval.md for guidance.

Setting	Default	Description
queryExpansionEnabled	false	Heuristic query expansion (no LLM calls)
queryExpansionMaxQueries	4	Max expanded queries including original
queryExpansionMinTokenLen	3	Minimum token length for expansion
rerankEnabled	false	LLM reranking pass over QMD/embedding results
rerankProvider	local	local only in v2.2
rerankMaxCandidates	20	Max candidates sent to reranker
rerankTimeoutMs	8000	Rerank timeout (ms)
rerankCacheEnabled	true	Cache reranks in-memory
rerankCacheTtlMs	3600000	Rerank cache TTL (ms)
feedbackEnabled	false	Enable memory_feedback tool and ranking bias
negativeExamplesEnabled	false	Track and penalize not-useful recalls
recencyWeight	0.2	Recency weight in retrieval ranking (0–1)
boostAccessCount	true	Boost frequently accessed memories in ranking
slowLogEnabled	false	Log slow operations
slowLogThresholdMs	30000	Threshold for slow log entries (ms)

v2.4 Context Retention

Setting	Default	Description
checkpointEnabled	true	Save a working-context checkpoint after each turn for recovery
checkpointTurns	15	Number of recent turns included in checkpoint context
transcriptEnabled	true	Save conversation transcripts to disk
transcriptRetentionDays	7	Days to retain saved transcripts
transcriptSkipChannelTypes	["cron"]	Channel types whose transcripts are not saved
transcriptRecallHours	12	Hours of transcript history to include in recall context
maxTranscriptTurns	50	Max turns of transcript context to inject
maxTranscriptTokens	1000	Token budget cap for transcript recall formatting
hourlySummariesEnabled	true	Generate hourly summaries of conversation activity
hourlySummaryCronAutoRegister	false	Auto-register hourly summary cron job on gateway start
hourlySummariesExtendedEnabled	false	Structured topics/decisions in hourly summaries
hourlySummariesIncludeToolStats	false	Include tool usage stats in summaries
conversationIndexEnabled	false	Index transcript chunks for semantic recall
conversationIndexBackend	qmd	Conversation index backend (qmd for QMD collections, faiss for the bundled local sidecar)
conversationIndexQmdCollection	openclaw-engram-conversations	QMD collection for conversation index
conversationIndexRetentionDays	30	Days of transcript chunks retained in the semantic conversation index
conversationIndexEmbedOnUpdate	false	Run qmd embed on each conversation-index update instead of batching embed runs separately
conversationIndexFaissScriptPath	(unset)	Optional absolute path to FAISS sidecar script
conversationIndexFaissPythonBin	(unset)	Optional Python executable for FAISS sidecar
conversationIndexFaissModelId	text-embedding-3-small	Embedding model id passed to the FAISS sidecar
conversationIndexFaissIndexDir	state/conversation-index/faiss	Relative FAISS artifact directory under memoryDir (index.faiss, metadata.jsonl, manifest.json)
conversationIndexFaissUpsertTimeoutMs	30000	Timeout for FAISS upsert operations
conversationIndexFaissSearchTimeoutMs	5000	Timeout for FAISS search operations
conversationIndexFaissHealthTimeoutMs	2000	Timeout for FAISS health checks; degraded health is fail-open
conversationIndexFaissMaxBatchSize	512	Max chunk batch size sent per FAISS upsert
conversationIndexFaissMaxSearchK	50	Max top-K allowed for FAISS search
conversationRecallTopK	3	Top-K relevant transcript chunks to inject
conversationRecallMaxChars	2500	Max characters of conversation context to inject
conversationRecallTimeoutMs	800	Timeout for conversation recall (ms)
conversationIndexMinUpdateIntervalMs	900000	Min interval between index updates

FAISS notes:

• conversation_index_update still writes chunk markdown under memoryDir/conversation-index/chunks/...; the FAISS backend additionally upserts those chunks into the local sidecar index.
• The sidecar health check reports degraded when Python dependencies or local artifacts are missing. Recall stays fail-open and skips semantic transcript injection instead of breaking hook execution.
• Sentence-transformers embeddings are opt-in via ENGRAM_FAISS_ENABLE_ST=1. Without that env var, the sidecar uses deterministic hash embeddings for low-friction local setups.

v9.1 Evaluation Harness Foundation

Setting	Default	Description
evalHarnessEnabled	false	Enable Engram's benchmark/evaluation harness bookkeeping
evalShadowModeEnabled	false	Record live recall decisions to the eval store without changing injected output
benchmarkBaselineSnapshotsEnabled	false	Enable versioned baseline snapshot artifacts for the latest completed benchmark runs
benchmarkDeltaReporterEnabled	false	Enable named-baseline delta reports against the current eval store
evalStoreDir	{memoryDir}/state/evals	Root directory for benchmark packs, run summaries, and shadow recall records
objectiveStateMemoryEnabled	false	Enable the objective-state memory foundation for normalized world/tool state snapshots
objectiveStateSnapshotWritesEnabled	false	Allow agent-end file/process/tool writers to persist objective-state snapshots into the store
objectiveStateRecallEnabled	false	Inject prompt-relevant objective-state snapshots into recall context
objectiveStateStoreDir	{memoryDir}/state/objective-state	Root directory for objective-state snapshot artifacts
causalTrajectoryMemoryEnabled	false	Enable the causal-trajectory memory foundation for typed goal-action-observation-outcome chains
causalTrajectoryStoreDir	{memoryDir}/state/causal-trajectories	Root directory for causal-trajectory records
causalTrajectoryRecallEnabled	false	Inject prompt-relevant causal trajectories into recall context
actionGraphRecallEnabled	false	Write action-conditioned causal-stage edges from typed trajectory records into the causal graph
trustZonesEnabled	false	Enable the trust-zone memory foundation and operator-facing promotion path for quarantine, working, and trusted records
quarantinePromotionEnabled	false	Allow explicit trust-zone promotions such as quarantine -> working and guarded working -> trusted
trustZoneStoreDir	{memoryDir}/state/trust-zones	Root directory for trust-zone records
trustZoneRecallEnabled	false	Inject prompt-relevant working and trusted trust-zone records into recall context
memoryPoisoningDefenseEnabled	false	Enable deterministic provenance trust scoring and corroboration requirements for risky trusted promotions
memoryRedTeamBenchEnabled	false	Enable typed memory-red-team benchmark packs and status accounting for poisoning-defense regression suites
harmonicRetrievalEnabled	false	Enable harmonic retrieval blending over abstraction nodes and cue anchors, including the dedicated recall section and harmonic-search diagnostics
abstractionAnchorsEnabled	false	Enable typed cue-anchor indexing for abstraction nodes and expose the anchor store through status tooling
abstractionNodeStoreDir	{memoryDir}/state/abstraction-nodes	Root directory for abstraction-node artifacts
verifiedRecallEnabled	false	Inject prompt-relevant memory boxes only when their cited source memories verify as non-archived episodes
semanticRulePromotionEnabled	false	Enable deterministic promotion of explicit IF ... THEN ... rules from verified episodic memories via openclaw engram semantic-rule-promote
semanticRuleVerificationEnabled	false	Verify promoted semantic rules against their cited source episodes at recall time and inject a dedicated Verified Rules section via openclaw engram semantic-rule-verify
creationMemoryEnabled	false	Enable the creation-memory foundation, including the typed work-product ledger and its operator-facing write/status commands
memoryUtilityLearningEnabled	false	Enable typed utility-learning telemetry storage, the offline learner commands openclaw engram utility-status, openclaw engram utility-record, openclaw engram utility-learning-status, and openclaw engram utility-learn, plus runtime loading of the persisted learner snapshot
promotionByOutcomeEnabled	false	Apply bounded learned utility weights to ranking heuristics and tier-migration thresholds when a learner snapshot is available
commitmentLedgerEnabled	false	Enable the explicit commitment ledger for promises, follow-ups, deadlines, and unfinished obligations
commitmentLifecycleEnabled	false	Enable commitment lifecycle transitions, stale tracking, and resolved-entry cleanup for the commitment ledger
commitmentStaleDays	14	Days before an open commitment without a due date is considered stale in lifecycle status
commitmentLedgerDir	{memoryDir}/state/commitment-ledger	Root directory for commitment ledger entries
resumeBundlesEnabled	false	Enable typed resume-bundle storage plus the operator-facing resume-bundle-status, resume-bundle-record, and resume-bundle-build commands
resumeBundleDir	{memoryDir}/state/resume-bundles	Root directory for resume bundles
workProductRecallEnabled	false	Inject prompt-relevant work-product ledger entries into recall and expose openclaw engram work-product-recall-search
workProductLedgerDir	{memoryDir}/state/work-product-ledger	Root directory for work-product ledger entries

Current foundation slice:

• openclaw engram benchmark-status scans benchmarks/**.json and runs/**.json, validates manifests/run summaries, and reports the latest completed run.
• When benchmarkBaselineSnapshotsEnabled is on, Engram also tracks typed baselines/*.json artifacts under the eval store and surfaces the latest stored baseline snapshot in openclaw engram benchmark-status.
• When both eval flags are on, live recall also writes shadow/YYYY-MM-DD/<trace-id>.json records with hashes, counts, chosen source, and recalled memory IDs.
• openclaw engram benchmark-validate <path> validates a manifest JSON file or a pack directory with a root manifest.json.
• openclaw engram benchmark-import <path> [--force] validates first, then imports into benchmarks/<benchmarkId>/.
• openclaw engram benchmark-baseline-snapshot --snapshot-id <id> captures a versioned baseline snapshot of the latest completed benchmark runs under baselines/<snapshotId>.json.
• openclaw engram benchmark-baseline-report --snapshot-id <id> compares the current eval store against a named stored baseline snapshot, emits both JSON and markdown summaries, and fails when pass rate, shared metrics, coverage, or eval artifact validity regress relative to that snapshot.
• The required GitHub eval-benchmark-gate workflow uses the committed fixture baseline snapshot at tests/fixtures/eval-ci/store/baselines/required-main.json as its stable PR-gating reference.
• openclaw engram benchmark-ci-gate --base <dir> --candidate <dir> compares two eval-store roots and fails when pass rate, shared metrics, or benchmark coverage regress.
• When objectiveStateRecallEnabled is on, Engram can inject a separate ## Objective State recall section sourced from the objective-state store.
• When causalTrajectoryMemoryEnabled is on, Engram can persist typed causal chains into a separate store for later graph/retrieval slices.
• When causalTrajectoryRecallEnabled is on, Engram can inject a separate ## Causal Trajectories recall section sourced from the causal-trajectory store.
• When actionGraphRecallEnabled is also on, each newly recorded causal trajectory emits deterministic goal -> action -> observation -> outcome -> follow_up edges into the causal graph without changing retrieval behavior yet.
• When trustZonesEnabled is on, Engram can persist provenance-bearing records into separate quarantine, working, and trusted storage tiers.
• When quarantinePromotionEnabled is also on, Engram exposes an explicit promotion path that blocks direct quarantine -> trusted jumps and requires anchored provenance before promoting risky working records into trusted.
• When trustZoneRecallEnabled is also on, Engram injects a separate ## Trust Zones recall section sourced from working and trusted trust-zone records while keeping quarantine records out of recall by default.
• When memoryPoisoningDefenseEnabled is also on, openclaw engram trust-zone-status reports deterministic provenance trust scores derived from source class plus sourceId / evidenceHash / sessionKey anchors so later poisoning defenses can build on explicit signals instead of hidden heuristics.
• With both memoryPoisoningDefenseEnabled and quarantinePromotionEnabled enabled, risky working -> trusted promotions now require at least one independent non-quarantine corroborating record with anchored provenance and overlapping entityRefs or tags.
• When memoryRedTeamBenchEnabled is on, benchmark manifests can also declare benchmarkType: "memory-red-team" plus attackClass and targetSurface, and openclaw engram benchmark-status reports red-team pack counts and unique attack metadata.
• When harmonicRetrievalEnabled is on, Engram can persist typed abstraction nodes into a separate abstraction-node store for later harmonic retrieval slices.
• When abstractionAnchorsEnabled is also on, Engram can persist cue-anchor index entries under {abstractionNodeStoreDir}/anchors for entities, files, tools, outcomes, constraints, and dates.
• When the harmonic retrieval section is enabled in the recall pipeline, Engram can inject a dedicated ## Harmonic Retrieval section that explains which abstraction nodes matched and which cue anchors contributed.
• Use openclaw engram abstraction-node-status to inspect node storage, openclaw engram cue-anchor-status to inspect anchor counts and invalid index records, and openclaw engram harmonic-search <query> to preview blended harmonic retrieval matches.
• When verifiedRecallEnabled is on, Engram can inject a separate ## Verified Episodes recall section sourced from recent memory boxes, but only when each surfaced box still cites at least one non-archived source memory whose memoryKind remains episode.
• Use openclaw engram verified-recall-search <query> to preview verified episodic recall matches, including verified memory counts, matched fields, and cited episodic memory IDs.
• When semanticRulePromotionEnabled is on, openclaw engram semantic-rule-promote --memory-id <id> can promote an explicit IF ... THEN ... rule from a non-archived episodic memory into a durable rule memory with lineage, sourceMemoryId, and duplicate suppression.
• When semanticRuleVerificationEnabled is on, Engram can inject a separate ## Verified Rules recall section sourced from promoted rule memories, but only when each surfaced rule still clears a provenance-aware effective-confidence threshold after re-checking its sourceMemoryId.
• When both creationMemoryEnabled and commitmentLedgerEnabled are on, Engram can persist explicit commitment ledger entries and expose them through openclaw engram commitment-status and openclaw engram commitment-record.
• When commitmentLifecycleEnabled is also on, Engram can transition commitment states with openclaw engram commitment-set-state, report overdue/stale/decay-eligible counts in openclaw engram commitment-status, and apply overdue-expiry plus resolved-entry cleanup through openclaw engram commitment-lifecycle-run.
• When both creationMemoryEnabled and resumeBundlesEnabled are on, Engram can persist explicit typed resume bundles, inspect them with openclaw engram resume-bundle-status, write manual shells with openclaw engram resume-bundle-record, and assemble bounded bundles from transcript recovery plus recent objective state, work products, and open commitments with openclaw engram resume-bundle-build.
• When creationMemoryEnabled is on, Engram can persist explicit work-product ledger entries and expose them through openclaw engram work-product-status and openclaw engram work-product-record.
• When both creationMemoryEnabled and workProductRecallEnabled are on, Engram can inject a separate ## Work Products recall section sourced from the typed work-product ledger and expose openclaw engram work-product-recall-search <query> for reuse previews.
• When memoryUtilityLearningEnabled is on, Engram can persist typed downstream utility telemetry for promotion and ranking decisions, inspect the resulting event ledger with openclaw engram utility-status, record explicit benchmark/operator utility observations through openclaw engram utility-record, and learn bounded offline promotion/ranking weights through openclaw engram utility-learn with the persisted learner snapshot visible in openclaw engram utility-learning-status.
• When promotionByOutcomeEnabled is also on and a learner snapshot exists, Engram applies bounded learned utility multipliers to ranking heuristic deltas and bounded promotion/demotion threshold nudges to tier migration without re-reading raw utility telemetry on the hot path.
• Use openclaw engram semantic-rule-verify <query> to preview verified semantic-rule matches, including verification status, effective confidence, and the cited source memory id.
• Future slices will add automated benchmark runners on top of this store and gate format.

v3.0 Namespaces

See namespaces.md.

Setting	Default	Description
namespacesEnabled	false	Enable multi-agent namespace isolation
defaultNamespace	default	Namespace for this agent's private memories
sharedNamespace	shared	Namespace for promoted shared memories
namespacePolicies	[]	Array of per-namespace read/write policy objects

v4.0 Shared Context

See shared-context.md.

Setting	Default	Description
sharedContextEnabled	false	Enable shared cross-agent context
sharedContextDir	(unset)	Directory for shared context files
sharedContextMaxInjectChars	4000	Max chars injected from shared context
sharedCrossSignalSemanticEnabled	false	Enable optional semantic overlap enhancement during daily curation
sharedCrossSignalSemanticTimeoutMs	4000	Timeout budget for semantic enhancement pass (fail-open on timeout)
sharedCrossSignalSemanticMaxCandidates	120	Max topic-token candidates considered by semantic enhancement

v5.0 Compounding

See compounding.md.

Setting	Default	Description
compoundingEnabled	false	Enable weekly synthesis and mistake learning
compoundingInjectEnabled	true	Inject compounding output when enabled

v6.0 Deduplication & Archival

Setting	Default	Description
factDeduplicationEnabled	true	Content-hash deduplication
semanticDedupEnabled	true	Write-time semantic similarity guard (issue #373) — embeds each candidate fact, queries the top-K nearest neighbors, and skips the write when cosine similarity ≥ semanticDedupThreshold. Fails open when the embedding backend is unavailable.
semanticDedupThreshold	0.92	Cosine similarity threshold in [0, 1] above which a candidate fact is treated as a near-duplicate and skipped.
semanticDedupCandidates	5	Number of nearest-neighbor candidates to compare against during the write-time semantic dedup check.
factArchivalEnabled	false	Archive old, low-value facts
factArchivalAgeDays	90	Minimum age to archive
factArchivalMaxImportance	0.3	Maximum importance to archive
factArchivalMaxAccessCount	2	Maximum access count to archive
factArchivalProtectedCategories	["commitment","preference","decision","principle","procedure"]	Never archived

Write-time semantic dedup (issue #373)

Exact content-hash dedup catches identical text but lets paraphrases through. The semantic guard runs in orchestrator.persistExtraction() after the hash miss and before any storage work:

1. Embed the candidate fact via the existing embedding-fallback pipeline.
2. Query the top semanticDedupCandidates (default 5) nearest neighbors.
3. If the best cosine similarity is ≥ semanticDedupThreshold (default 0.92), drop the candidate, bump the existing dedupedCount metric, and log a debug line naming the colliding neighbor id and score.
4. On any lookup error or missing backend, the guard fails open so writes are never blocked by an embedding outage.

The guard shares its decision function with the CLI remnic dedup tooling via packages/remnic-core/src/dedup/semantic.ts, so there is a single source of truth for similarity logic across read-time and write-time code paths.

v8.2 Graph Recall Activation

Setting	Default	Description
multiGraphMemoryEnabled	false	Enable graph storage/traversal substrate
graphRecallEnabled	false	Enable planner graph_mode expansion
graphExpandedIntentEnabled	true	Escalate broader causal/timeline prompts into graph_mode
graphAssistInFullModeEnabled	true	Run bounded graph expansion during full recall mode
graphAssistShadowEvalEnabled	false	In full mode, run graph assist as shadow-eval (compute + snapshot + telemetry, no injection change)
graphAssistMinSeedResults	3	Minimum seed recalls required for full-mode graph assist
graphWriteSessionAdjacencyEnabled	true	Write fallback time edges between consecutive extracted memories
entityGraphEnabled	true	Enable entity co-reference edges
timeGraphEnabled	true	Enable temporal sequence edges
causalGraphEnabled	true	Enable causal phrase edges
maxGraphTraversalSteps	3	Max spreading-activation BFS hops
graphActivationDecay	0.7	Per-hop decay factor
graphTraversalConfidenceFloor	0.2	Minimum edge confidence required for traversal (issue #681 PR 3/3). Edges below this floor are pruned. Legacy edges without confidence are treated as 1.0. Range [0, 1].
graphTraversalPageRankIterations	8	PageRank-style refinement iterations applied on top of BFS spreading-activation scores (issue #681 PR 3/3). Set to 0 to disable refinement.
graphEdgeDecayEnabled	false	Enable the periodic graph-edge confidence decay maintenance job (issue #681 PR 2/3). When false all edges retain their initial confidence indefinitely.
graphEdgeDecayCadenceMs	604800000	How often the decay job runs, in milliseconds (default 7 days). Minimum enforced at 60000 ms.
graphEdgeDecayWindowMs	7776000000	Length of one decay window, in milliseconds (default 90 days). One window of inactivity costs graphEdgeDecayPerWindow confidence. Minimum enforced at 60000 ms.
graphEdgeDecayPerWindow	0.1	Fraction of confidence lost per elapsed decay window. Range [0, 1].
graphEdgeDecayFloor	0.1	Minimum confidence an edge can decay to; the job will not reduce confidence below this value. Range [0, 1]. Set to 0 to allow full decay to zero.
graphExpansionActivationWeight	0.65	Blend weight for graph activation vs seed QMD score (0-1)
graphExpansionBlendMin	0.05	Lower clamp bound for blended graph-expanded scores (0-1)
graphExpansionBlendMax	0.95	Upper clamp bound for blended graph-expanded scores (0-1)

File Hygiene

Setting	Default	Description
fileHygiene.enabled	false	Enable file hygiene features
fileHygiene.lintEnabled	true	Warn on oversized workspace files (when hygiene is enabled)
fileHygiene.lintPaths	["IDENTITY.md","MEMORY.md"]	Files to monitor (relative to workspaceDir)
fileHygiene.lintBudgetBytes	20000	Budget threshold for warnings
fileHygiene.lintWarnRatio	0.8	Warn at this fraction of budget
fileHygiene.rotateEnabled	false	Rotate oversized files into archive
fileHygiene.rotatePaths	["IDENTITY.md"]	Files to rotate
fileHygiene.rotateMaxBytes	18000	Max size before rotation
fileHygiene.rotateKeepTailChars	2000	Chars to keep as tail excerpt after rotation
fileHygiene.archiveDir	.engram-archive	Archive directory name
fileHygiene.runMinIntervalMs	300000	Min interval between hygiene runs
fileHygiene.warningsLogEnabled	false	Write human-readable hygiene warnings into the workspace instead of logging only to the gateway log
fileHygiene.warningsLogPath	hygiene/warnings.md	Workspace-relative warnings log path used when warning logging is enabled
fileHygiene.indexEnabled	false	Maintain an optional operator-facing workspace index file during hygiene passes
fileHygiene.indexPath	ENGRAM_INDEX.md	Workspace-relative path for the optional generated index file

Access Tracking

Setting	Default	Description
accessTrackingEnabled	true	Track access frequency per memory
boostAccessCount	true	Boost frequently accessed memories in ranking

Memory Linking

Setting	Default	Description
memoryLinkingEnabled	false	LLM-suggested semantic links between memories

Summarization

Setting	Default	Description
summarizationEnabled	false	Summarize old memories when count exceeds threshold
summarizationTriggerCount	1000	Memory count that triggers summarization

Extraction Judge (issue #376)

Setting	Default	Description
extractionJudgeEnabled	false	Enable the LLM-as-judge post-extraction durability filter
extractionJudgeModel	""	Model override for judge; empty = use configured local model
extractionJudgeBatchSize	20	Max candidates per LLM batch call
extractionJudgeShadow	false	Shadow mode: log verdicts without filtering

Semantic Chunking (issue #368)

Setting	Default	Description
semanticChunkingEnabled	false	Enable topic-boundary chunking via sentence embeddings
semanticChunkingConfig	(see below)	Sub-object with chunking parameters

semanticChunkingConfig keys

Key	Type	Default	Description
targetTokens	number	200	Target token count per chunk
minTokens	number	100	Minimum token count per chunk
maxTokens	number	400	Maximum token count per chunk
smoothingWindowSize	number	3	Sliding window size for similarity smoothing
boundaryThresholdStdDevs	number	1.0	Standard deviations below mean similarity to trigger a boundary
embeddingBatchSize	number	32	Batch size for sentence embedding calls
fallbackToRecursive	boolean	true	Fall back to recursive character chunking when embeddings are unavailable

Page Versioning (issue #371)

Setting	Default	Description
versioningEnabled	false	Enable page-level versioning
versioningMaxPerPage	50	Max snapshots per page (0 = unlimited)
versioningSidecarDir	".versions"	Override sidecar directory path (.versions/ relative to memoryDir when unset)

Citations (issue #379)

Setting	Default	Description
emitLegacyTools	true	Advertise legacy engram_*/engram.* MCP tool aliases alongside canonical remnic_* names. Set false to halve the advertised tools/list surface (only remnic_*); tools stay callable under both names. Also settable via REMNIC_EMIT_LEGACY_TOOLS. (issue #1427)
citationsEnabled	false	Emit oai-mem-citation blocks in recall responses
citationsAutoDetect	true	Auto-detect Codex citation context

MECE Taxonomy (issue #366)

Setting	Default	Description
taxonomyEnabled	false	Enable the MECE knowledge directory
taxonomyAutoGenResolver	true	Auto-regenerate RESOLVER.md when taxonomy changes

Enrichment Pipeline (issue #365)

Setting	Default	Description
enrichmentEnabled	false	Enable external entity enrichment pipeline
enrichmentAutoOnCreate	false	Auto-enrich newly created entities
enrichmentMaxCandidatesPerEntity	20	Max enrichment candidates per entity per run

Binary Lifecycle (issue #367)

Setting	Default	Description
binaryLifecycleEnabled	false	Enable binary file lifecycle management
binaryLifecycleGracePeriodDays	7	Days before local cleanup of mirrored files
binaryLifecycleBackendType	"none"	Storage backend: "none", "filesystem", "s3"
binaryLifecycleBackendPath	""	Base path for filesystem backend

Peer registry (issue #679)

Manages the multi-peer schema (self, human, agent, integration). Full narrative: peers.md.

Profile reasoner

The async profile reasoner derives structured profile.md fields from interaction-log signals. It runs inside the Dreams REM phase and is disabled by default.

Setting	Default	Description
peerProfileReasonerEnabled	false	Master gate for the async peer profile reasoner. Set to true to opt in. Least-privileged default per CLAUDE.md rules #30/#48.
peerProfileReasonerModel	"auto"	Model routing alias used for reasoner LLM calls. "auto" uses the platform routing default; operators can pin to a specific model identifier.
peerProfileReasonerMinInteractions	5	Minimum interaction-log entries required before the reasoner processes a peer. 0 disables the minimum (process every peer regardless of log depth).
peerProfileReasonerMaxFieldsPerRun	8	Maximum profile fields the reasoner may write per peer per run. 0 disables field writes while still allowing the reasoner to run its analysis.

Recall injection

Injects a ## Peer Profile section from the peer registered for the active session. Disabled by default.

Setting	Default	Description
peerProfileRecallEnabled	false	Gate for peer-profile injection into recall context. Set to true to inject the active session's peer profile into every recall.
peerProfileRecallMaxFields	5	Maximum number of profile fields injected per recall. Fields are selected by most-recently-updated provenance timestamp. 0 disables injection even when peerProfileRecallEnabled is true.

Procedural memory (issue #519)

Stored as category: procedure markdown under memoryDir/procedures/. Narrative overview: procedural-memory.md.

Setting	Default	Description
procedural.enabled	true	Master gate: default-on since issue #567 PR 4/5 (previously false). Set to false (or any of "0", "no", "off") to opt out of procedure extraction writes, task-initiation procedure recall injection, and trajectory mining side effects.
procedural.minOccurrences	3	Minimum cluster size for a candidate; clusters smaller than this are skipped. 0 disables procedural mining (runProcedureMining returns immediately with skippedReason: "minOccurrences_zero").
procedural.successFloor	0.75	Minimum trajectory success rate in [0, 1] for miner eligibility. Raised from 0.7 in issue #567 PR 3/5.
procedural.autoPromoteOccurrences	8	When auto-promotion is on, occurrences before pending_review → active.
procedural.autoPromoteEnabled	false	Allow automatic promotion of miner candidates that meet thresholds.
procedural.lookbackDays	14	Trajectory lookback window for mining (days). Lowered from 30 in issue #567 PR 3/5.
procedural.proceduralMiningCronAutoRegister	false	When true, installer may register the nightly procedural mining cron entry.
procedural.recallMaxProcedures	2	Max procedure previews injected on task-initiation recall (1–10). Lowered from 3 in issue #567 PR 3/5 so procedural injection does not crowd other recall sections.

Pattern reinforcement (issue #687)

Cross-session pattern detection: clusters memories by normalized content, reinforces recurring primitives with reinforcement_count + last_reinforced_at, and optionally boosts their recall score. Narrative overview: pattern-reinforcement.md.

Setting	Default	Description
patternReinforcementEnabled	false	Master gate. Set to true to enable the maintenance job that detects and reinforces recurring memory patterns across sessions. Default false (opt-in).
patternReinforcementCadenceMs	604800000	Minimum milliseconds between pattern-reinforcement runs (default 7 days). Set to 0 to disable cadence gating and allow the job to run on every MCP/cron invocation.
patternReinforcementMinCount	3	Minimum cluster size before a canonical memory is promoted and reinforced. Clamped to [2, 1000]; clusters of 1 are degenerate.
patternReinforcementCategories	["preference", "fact", "decision"]	Memory categories the job considers. Set to [] to process no categories. Procedure memories are intentionally excluded from the default list to avoid interference with the procedural miner.
reinforcementRecallBoostEnabled	false	When true, memories with reinforcement_count > 0 receive an additive score boost during recall. Default false (opt-in). Requires patternReinforcementEnabled: true upstream to populate reinforcement counts.
reinforcementRecallBoostWeight	0.05	Per-unit score bonus applied per reinforcement_count. Raw boost is weight × reinforcement_count, then clipped at reinforcementRecallBoostMax. Range [0, 1].
reinforcementRecallBoostMax	0.3	Maximum additive reinforcement boost per recall result. Range [0, 1]. Raw boost formula: min(reinforcementRecallBoostMax, reinforcementRecallBoostWeight × reinforcement_count).

Codex Marketplace (issue #418)

Setting	Default	Description
codexMarketplaceEnabled	true	Enable Codex marketplace installation support

Memory Extensions (issue #382)

Setting	Default	Description
memoryExtensionsEnabled	true	Enable third-party memory extension discovery

Cross-namespace Query Budget (issue #565)

Per-principal sliding-window rate limiter for cross-namespace recall queries. When enabled, principals issuing bursts of recalls against namespaces other than their own are throttled: soft limit emits a warning, hard limit denies the query. See Threat model.

Setting	Default	Description
recallCrossNamespaceBudgetEnabled	false	Enable per-principal cross-namespace recall budget
recallCrossNamespaceBudgetWindowMs	60000	Sliding window duration in milliseconds
recallCrossNamespaceBudgetSoftLimit	10	Queries per window that trigger a warning (still allowed)
recallCrossNamespaceBudgetHardLimit	30	Queries per window that trigger a denial

Recall Audit Anomaly Detection (issue #565)

Anomaly detection on the recall audit trail. Flags suspicious query patterns (repeat queries, namespace walks, high-cardinality entity probes, rapid-fire) in recall responses. See Threat model.

Setting	Default	Description
recallAuditAnomalyDetectionEnabled	false	Enable anomaly detection on recall audit trail
recallAuditAnomalyWindowMs	300000	Sliding window for anomaly detectors (5 min)
recallAuditAnomalyRepeatQueryLimit	5	Max identical queries before repeat-query flag
recallAuditAnomalyNamespaceWalkLimit	3	Max distinct namespaces before namespace-walk flag
recallAuditAnomalyHighCardinalityLimit	50	Max candidate memory IDs in a single recall response before high-cardinality flag
recallAuditAnomalyRapidFireLimit	30	Max queries in window before rapid-fire flag
memoryExtensionsRoot	""	Override memory extensions root directory

Schema-Complete Default and Recommended Settings

This appendix is flattened from the runtime config schema and the live parseConfig({}) defaults so the page stays complete even when newer or advanced settings have not yet been expanded in the narrative sections above. Unless noted otherwise, the recommended value matches the shipped default.

Setting	Default	Recommended
openaiApiKey	(env fallback in plugin mode)	unset when modelSource is gateway; set false for local-only plugin mode; otherwise explicit key or OPENAI_API_KEY env fallback
openaiBaseUrl	(unset)	(unset)
model	gpt-5.5	gpt-5.5
reasoningEffort	low	low
triggerMode	smart	smart
bufferMaxTurns	5	5
bufferMaxMinutes	15	15
consolidateEveryN	3	3
highSignalPatterns	[]	[]
maxMemoryTokens	2000	2000
memoryOsPreset	(unset)	balanced
qmdEnabled	true	true
qmdCollection	openclaw-engram	openclaw-engram
qmdMaxResults	8	8
qmdColdTierEnabled	false	false unless you are actively tiering hot/cold QMD collections
qmdColdCollection	openclaw-engram-cold	openclaw-engram-cold
qmdColdMaxResults	8	8
qmdTierMigrationEnabled	false	false unless hot/cold QMD tiering is enabled
qmdTierDemotionMinAgeDays	14	14
qmdTierDemotionValueThreshold	0.35	0.35
qmdTierPromotionValueThreshold	0.7	0.7
qmdTierParityGraphEnabled	true	true
qmdTierParityHiMemEnabled	true	true
qmdTierAutoBackfillEnabled	false	false
embeddingFallbackEnabled	true	true
embeddingFallbackProvider	auto	auto
qmdPath	(unset)	(unset)
memoryDir	~/.openclaw/workspace/memory/local	~/.openclaw/workspace/memory/local
debug	false	false
identityEnabled	true	true
identityContinuityEnabled	false	false
identityInjectionMode	recovery_only	recovery_only
identityMaxInjectChars	1200	1200
continuityIncidentLoggingEnabled	false	false
continuityAuditEnabled	false	false
sessionObserverEnabled	false	false until you are ready for heartbeat-triggered extraction
sessionObserverDebounceMs	120000	120000
sessionObserverBands	[{"maxBytes":50000,"triggerDeltaBytes":4800,"triggerDeltaTokens":1200},{"maxBytes":200000,"triggerDeltaBytes":9600,"triggerDeltaTokens":2400},{"maxBytes":1000000000,"triggerDeltaBytes":19200,"triggerDeltaTokens":4800}]	[{"maxBytes":50000,"triggerDeltaBytes":4800,"triggerDeltaTokens":1200},{"maxBytes":200000,"triggerDeltaBytes":9600,"triggerDeltaTokens":2400},{"maxBytes":1000000000,"triggerDeltaBytes":19200,"triggerDeltaTokens":4800}]
sessionObserverBands[].maxBytes	50000	50000
sessionObserverBands[].triggerDeltaBytes	4800	4800
sessionObserverBands[].triggerDeltaTokens	1200	1200
injectQuestions	false	false
commitmentDecayDays	90	90
workspaceDir	~/.openclaw/workspace	~/.openclaw/workspace
fileHygiene	(unset)	(unset)
fileHygiene.enabled	false	true
fileHygiene.lintEnabled	true	true
fileHygiene.lintBudgetBytes	20000	20000
fileHygiene.lintWarnRatio	0.8	0.8
fileHygiene.lintPaths	["IDENTITY.md","MEMORY.md"]	["IDENTITY.md","MEMORY.md"]
fileHygiene.rotateEnabled	false	false
fileHygiene.rotateMaxBytes	18000	18000
fileHygiene.rotateKeepTailChars	2000	2000
fileHygiene.rotatePaths	["IDENTITY.md"]	["IDENTITY.md"]
fileHygiene.archiveDir	.engram-archive	.engram-archive
fileHygiene.runMinIntervalMs	300000	300000
fileHygiene.warningsLogEnabled	false	false
fileHygiene.warningsLogPath	hygiene/warnings.md	hygiene/warnings.md
fileHygiene.indexEnabled	false	false
fileHygiene.indexPath	ENGRAM_INDEX.md	ENGRAM_INDEX.md
nativeKnowledge	(unset)	(unset)
nativeKnowledge.enabled	false	true when workspace bootstrap docs exist
nativeKnowledge.includeFiles	["IDENTITY.md","MEMORY.md"]	["IDENTITY.md","MEMORY.md"]
nativeKnowledge.maxChunkChars	900	900
nativeKnowledge.maxResults	4	4
nativeKnowledge.maxChars	2400	2400
nativeKnowledge.stateDir	state/native-knowledge	state/native-knowledge
nativeKnowledge.openclawWorkspace	(unset)	(unset)
nativeKnowledge.openclawWorkspace.enabled	false	true when you want handoffs/daily summaries in recall
nativeKnowledge.openclawWorkspace.bootstrapFiles	["IDENTITY.md","MEMORY.md","USER.md"]	["IDENTITY.md","MEMORY.md","USER.md"]
nativeKnowledge.openclawWorkspace.handoffGlobs	["**/*handoff*.md","handoffs/**/*.md"]	["**/*handoff*.md","handoffs/**/*.md"]
nativeKnowledge.openclawWorkspace.dailySummaryGlobs	["**/*daily*summary*.md","summaries/**/*.md"]	["**/*daily*summary*.md","summaries/**/*.md"]
nativeKnowledge.openclawWorkspace.automationNoteGlobs	[]	[]
nativeKnowledge.openclawWorkspace.workspaceDocGlobs	[]	[]
nativeKnowledge.openclawWorkspace.excludeGlobs	[]	[]
nativeKnowledge.openclawWorkspace.sharedSafeGlobs	[]	[]
nativeKnowledge.obsidianVaults	[]	[]
nativeKnowledge.obsidianVaults[].id	(unset)	set explicitly for every configured vault
nativeKnowledge.obsidianVaults[].rootDir	(unset)	(unset)
nativeKnowledge.obsidianVaults[].includeGlobs	["**/*.md"]	["**/*.md"]
nativeKnowledge.obsidianVaults[].excludeGlobs	[".obsidian/**","**/*.canvas","**/*.png","**/*.jpg","**/*.jpeg","**/*.gif","**/*.pdf"]	[".obsidian/**","**/*.canvas","**/*.png","**/*.jpg","**/*.jpeg","**/*.gif","**/*.pdf"]
nativeKnowledge.obsidianVaults[].namespace	(unset)	(unset)
nativeKnowledge.obsidianVaults[].privacyClass	(unset)	(unset)
nativeKnowledge.obsidianVaults[].folderRules	[]	[]
nativeKnowledge.obsidianVaults[].folderRules[].pathPrefix	(unset)	(unset)
nativeKnowledge.obsidianVaults[].folderRules[].namespace	(unset)	(unset)
nativeKnowledge.obsidianVaults[].folderRules[].privacyClass	(unset)	(unset)
nativeKnowledge.obsidianVaults[].dailyNotePatterns	["YYYY-MM-DD"]	["YYYY-MM-DD"]
nativeKnowledge.obsidianVaults[].materializeBacklinks	false	false
agentAccessHttp	{"enabled":false,"host":"127.0.0.1","port":4318,"maxBodyBytes":131072}	{"enabled":false,"host":"127.0.0.1","port":4318,"maxBodyBytes":131072}
agentAccessHttp.enabled	false	false unless you need the local HTTP bridge
agentAccessHttp.host	127.0.0.1	127.0.0.1
agentAccessHttp.port	4318	4318
agentAccessHttp.authToken	(unset)	set explicitly whenever agentAccessHttp.enabled=true
agentAccessHttp.maxBodyBytes	131072	131072
accessTrackingEnabled	true	true
accessTrackingBufferMaxSize	100	100
recencyWeight	0.2	0.2
boostAccessCount	true	true
recordEmptyRecallImpressions	false	false
recallPlannerEnabled	true	true
recallPlannerLlmEnabled	false	false
recallPlannerModel	gpt-5.5	gpt-5.5
recallPlannerTimeoutMs	1500	1500
recallPlannerUseResponsesApi	true	true
recallPlannerMaxPromptChars	4000	4000
recallPlannerMaxMemoryHints	24	24
recallPlannerShadowMode	false	false
recallPlannerTelemetryEnabled	true	true
recallPlannerMaxQmdResultsMinimal	4	4
recallPlannerMaxQmdResultsFull	8	8
intentRoutingEnabled	false	false
intentRoutingBoost	0.12	0.12
verbatimArtifactsEnabled	false	true
verbatimArtifactsMinConfidence	0.8	0.8
verbatimArtifactsMaxRecall	5	5
verbatimArtifactCategories	["decision","correction","principle","commitment"]	["decision","correction","principle","commitment"]
memoryBoxesEnabled	false	false
boxTopicShiftThreshold	0.35	0.35
boxTimeGapMs	1800000	1800000
boxMaxMemories	50	50
traceWeaverEnabled	false	false
traceWeaverLookbackDays	7	7
traceWeaverOverlapThreshold	0.4	0.4
boxRecallDays	3	3
episodeNoteModeEnabled	false	false
queryAwareIndexingEnabled	false	true
queryAwareIndexingMaxCandidates	200	200
temporalIndexWindowDays	30	30
temporalIndexMaxEntries	5000	5000
temporalBoostRecentDays	7	7
temporalBoostScore	0.15	0.15
temporalDecayEnabled	true	true
tagMemoryEnabled	false	false
tagMaxPerMemory	5	5
tagIndexMaxEntries	10000	10000
tagRecallBoost	0.15	0.15
tagRecallMaxMatches	10	10
multiGraphMemoryEnabled	false	false
graphRecallEnabled	false	false
graphRecallMaxExpansions	3	3
graphRecallMaxPerSeed	5	5
graphRecallMinEdgeWeight	0.1	0.1
graphRecallShadowEnabled	false	false
graphRecallSnapshotEnabled	false	false
graphRecallShadowSampleRate	0.1	0.1
graphRecallExplainToolEnabled	false	false
graphRecallStoreColdMirror	false	false
graphRecallColdMirrorCollection	(unset)	(unset)
graphRecallColdMirrorMinAgeDays	7	7
graphRecallUseEntityPriors	false	false
graphRecallEntityPriorBoost	0.2	0.2
graphRecallPreferHubSeeds	false	false
graphRecallHubBias	0.3	0.3
graphRecallRecencyHalfLifeDays	30	30
graphRecallDampingFactor	0.85	0.85
graphRecallMaxSeedNodes	10	10
graphRecallMaxExpandedNodes	30	30
graphRecallMaxTrailPerNode	5	5
graphRecallMinSeedScore	0.3	0.3
graphRecallExpansionScoreThreshold	0.2	0.2
graphRecallExplainMaxPaths	3	3
graphRecallExplainMaxChars	500	500
graphRecallExplainEdgeLimit	5	5
graphRecallExplainEnabled	false	false
graphRecallEntityHintsEnabled	false	false
graphRecallEntityHintMax	3	3
graphRecallEntityHintMaxChars	200	200
graphRecallSnapshotDir	~/.openclaw/workspace/memory/local/state/graph	~/.openclaw/workspace/memory/local/state/graph
graphRecallEnableTrace	false	false
graphRecallEnableDebug	false	false
graphExpandedIntentEnabled	true	true
graphAssistInFullModeEnabled	true	true
graphAssistShadowEvalEnabled	false	false
graphAssistMinSeedResults	3	3
entityGraphEnabled	true	true
timeGraphEnabled	true	true
graphWriteSessionAdjacencyEnabled	true	true
causalGraphEnabled	true	true
maxGraphTraversalSteps	3	3
graphActivationDecay	0.7	0.7
graphExpansionActivationWeight	0.65	0.65
graphExpansionBlendMin	0.05	0.05
graphExpansionBlendMax	0.95	0.95
maxEntityGraphEdgesPerMemory	10	10
delinearizeEnabled	true	true
recallConfidenceGateEnabled	false	false
recallConfidenceGateThreshold	0.12	0.12
causalRuleExtractionEnabled	false	false
memoryReconstructionEnabled	false	false
memoryReconstructionMaxExpansions	3	3
graphLateralInhibitionEnabled	true	true
graphLateralInhibitionBeta	0.15	0.15
graphLateralInhibitionTopM	7	7
temporalMemoryTreeEnabled	false	false
tmtHourlyMinMemories	3	3
tmtSummaryMaxTokens	300	300
queryExpansionEnabled	false	false
queryExpansionMaxQueries	4	4
queryExpansionMinTokenLen	3	3
rerankEnabled	false	true
rerankProvider	local	local
rerankMaxCandidates	20	20
rerankTimeoutMs	8000	8000
rerankCacheEnabled	true	true
rerankCacheTtlMs	3600000	3600000
feedbackEnabled	false	false until operators are actively curating recall quality
negativeExamplesEnabled	false	false until operators are actively curating negative examples
negativeExamplesPenaltyPerHit	0.05	0.05
negativeExamplesPenaltyCap	0.25	0.25
chunkingEnabled	false	false
chunkingTargetTokens	200	200
chunkingMinTokens	150	150
chunkingOverlapSentences	2	2
contradictionDetectionEnabled	false	false
contradictionSimilarityThreshold	0.7	0.7
contradictionMinConfidence	0.9	0.9
contradictionAutoResolve	true	true
memoryLinkingEnabled	false	false
threadingEnabled	false	false
threadingGapMinutes	30	30
summarizationEnabled	false	false
summarizationTriggerCount	1000	1000
summarizationRecentToKeep	300	300
summarizationImportanceThreshold	0.3	0.3
summarizationProtectedTags	["commitment","preference","decision","principle"]	["commitment","preference","decision","principle"]
topicExtractionEnabled	true	true
topicExtractionTopN	50	50
transcriptEnabled	true	true
captureMode	implicit	implicit
transcriptRetentionDays	7	7
transcriptSkipChannelTypes	["cron"]	["cron"]
transcriptRecallHours	12	12
maxTranscriptTurns	50	50
maxTranscriptTokens	1000	1000
checkpointEnabled	true	true
checkpointTurns	15	15
compactionResetEnabled	false	false
hourlySummariesEnabled	true	true
summaryRecallHours	24	24
maxSummaryCount	6	6
summaryModel	gpt-5.5	gpt-5.5
localLlmEnabled	false	false unless you have a healthy compatible endpoint
localLlmUrl	http://localhost:1234/v1	http://localhost:1234/v1
localLlmModel	local-model	local-model
localLlmApiKey	(unset)	(unset)
localLlmHeaders	(unset)	(unset)
localLlmAuthHeader	true	true
localLlmFallback	true	true
localLlmHomeDir	(unset)	(unset)
localLmsCliPath	(unset)	(unset)
localLmsBinDir	(unset)	(unset)
localLlmTimeoutMs	180000	180000
slowLogEnabled	false	false
slowLogThresholdMs	30000	30000
traceRecallContent	false	false
extractionDedupeEnabled	true	true
extractionDedupeWindowMs	300000	300000
extractionMinChars	40	40
extractionMinUserTurns	1	1
extractionMaxTurnChars	4000	4000
extractionMaxFactsPerRun	12	12
extractionMaxEntitiesPerRun	6	6
extractionMaxQuestionsPerRun	3	3
extractionMaxProfileUpdatesPerRun	4	4
consolidationRequireNonZeroExtraction	true	true
consolidationMinIntervalMs	600000	600000
qmdMaintenanceEnabled	true	true
qmdMaintenanceDebounceMs	30000	30000
maintenance.namespaceFanoutEnabled / maintenanceNamespaceFanoutEnabled	true	When namespaces are enabled, allow background jobs to use the namespace catalog to discover dynamic project/team namespaces.
maintenance.maxNamespacesPerCycle / maintenanceMaxNamespacesPerCycle	20	Deterministic cap for each namespace-aware maintenance cycle. Default/shared/configured namespaces keep priority.
maintenance.includeProjectNamespaces / maintenanceIncludeProjectNamespaces	true	Include project namespaces discovered from the catalog when their live roots contain memory data.
maintenance.includeBranchNamespaces / maintenanceIncludeBranchNamespaces	false	Include branch namespaces in fanout. Off by default to avoid runaway branch maintenance.
maintenance.includeTeamProjectNamespaces / maintenanceIncludeTeamProjectNamespaces	true	Include team-project namespaces discovered from trusted scope/profile writes.
maintenance.namespaceLockStaleMs / maintenanceNamespaceLockStaleMs	600000	Stale threshold for per-job/per-namespace maintenance locks under state/maintenance-locks/.
qmdAutoEmbedEnabled	false	false
qmdEmbedMinIntervalMs	3600000	3600000
qmdUpdateTimeoutMs	90000	90000
qmdUpdateMinIntervalMs	900000	900000
localLlmRetry5xxCount	1	1
localLlmRetryBackoffMs	400	400
localLlm400TripThreshold	5	5
localLlm400CooldownMs	120000	120000
localLlmMaxContext	(unset)	(unset)
localLlmFastEnabled	false	false unless you have a separate fast local tier
localLlmFastModel	""	""
localLlmFastUrl	http://localhost:1234/v1	http://localhost:1234/v1
localLlmFastTimeoutMs	15000	15000
localLlmDisableThinking	true	true (skip reasoning tokens on structured extraction); set false if you want thinking on narrative tasks
hourlySummaryCronAutoRegister	false	false
hourlySummariesExtendedEnabled	false	false unless structured hourly summaries are useful
hourlySummariesIncludeToolStats	false	false
hourlySummariesIncludeSystemMessages	false	false
hourlySummariesMaxTurnsPerRun	200	200
conversationIndexEnabled	false	false unless you want transcript semantic recall
conversationIndexBackend	qmd	qmd
conversationIndexQmdCollection	openclaw-engram-conversations	openclaw-engram-conversations
conversationIndexRetentionDays	30	30
conversationIndexMinUpdateIntervalMs	900000	900000
conversationIndexEmbedOnUpdate	false	false
conversationIndexFaissScriptPath	""	""
conversationIndexFaissPythonBin	""	""
conversationIndexFaissModelId	text-embedding-3-small	text-embedding-3-small
conversationIndexFaissIndexDir	state/conversation-index/faiss	state/conversation-index/faiss
conversationIndexFaissUpsertTimeoutMs	30000	30000
conversationIndexFaissSearchTimeoutMs	5000	5000
conversationIndexFaissHealthTimeoutMs	2000	2000
conversationIndexFaissMaxBatchSize	512	512
conversationIndexFaissMaxSearchK	50	50
conversationRecallTopK	3	4
conversationRecallMaxChars	2500	2000
conversationRecallTimeoutMs	800	800
evalHarnessEnabled	false	false
evalShadowModeEnabled	false	false
benchmarkBaselineSnapshotsEnabled	false	false
benchmarkStoredBaselineEnabled	false	false
benchmarkDeltaReporterEnabled	false	false
evalStoreDir	~/.openclaw/workspace/memory/local/state/evals	~/.openclaw/workspace/memory/local/state/evals
objectiveStateMemoryEnabled	false	false
objectiveStateSnapshotWritesEnabled	false	false
objectiveStateRecallEnabled	false	false
objectiveStateStoreDir	~/.openclaw/workspace/memory/local/state/objective-state	~/.openclaw/workspace/memory/local/state/objective-state
causalTrajectoryMemoryEnabled	false	false
causalTrajectoryStoreDir	~/.openclaw/workspace/memory/local/state/causal-trajectories	~/.openclaw/workspace/memory/local/state/causal-trajectories
causalTrajectoryRecallEnabled	false	false
trustZonesEnabled	false	false
quarantinePromotionEnabled	false	false
trustZoneStoreDir	~/.openclaw/workspace/memory/local/state/trust-zones	~/.openclaw/workspace/memory/local/state/trust-zones
trustZoneRecallEnabled	false	false
memoryPoisoningDefenseEnabled	false	false
memoryRedTeamBenchEnabled	false	false
harmonicRetrievalEnabled	false	false
abstractionAnchorsEnabled	false	false
abstractionNodeStoreDir	~/.openclaw/workspace/memory/local/state/abstraction-nodes	~/.openclaw/workspace/memory/local/state/abstraction-nodes
verifiedRecallEnabled	false	false
semanticRulePromotionEnabled	false	false
semanticRuleVerificationEnabled	false	false
creationMemoryEnabled	false	false
memoryUtilityLearningEnabled	false	false
promotionByOutcomeEnabled	false	false
commitmentLedgerEnabled	false	false
commitmentLifecycleEnabled	false	false
commitmentStaleDays	14	14
commitmentLedgerDir	~/.openclaw/workspace/memory/local/state/commitment-ledger	~/.openclaw/workspace/memory/local/state/commitment-ledger
resumeBundlesEnabled	false	false
resumeBundleDir	~/.openclaw/workspace/memory/local/state/resume-bundles	~/.openclaw/workspace/memory/local/state/resume-bundles
workProductRecallEnabled	false	false
workProductLedgerDir	~/.openclaw/workspace/memory/local/state/work-product-ledger	~/.openclaw/workspace/memory/local/state/work-product-ledger
workTasksEnabled	false	false
workProjectsEnabled	false	false
workTasksDir	~/.openclaw/workspace/memory/local/work/tasks	~/.openclaw/workspace/memory/local/work/tasks
workProjectsDir	~/.openclaw/workspace/memory/local/work/projects	~/.openclaw/workspace/memory/local/work/projects
workIndexEnabled	false	false
workIndexDir	~/.openclaw/workspace/memory/local/work/index	~/.openclaw/workspace/memory/local/work/index
workTaskIndexEnabled	false	false
workProjectIndexEnabled	false	false
workIndexAutoRebuildEnabled	false	false
workIndexAutoRebuildDebounceMs	1000	1000
actionGraphRecallEnabled	false	false
namespacesEnabled	false	false
defaultNamespace	default	default
sharedNamespace	shared	shared
principalFromSessionKeyMode	map	map
principalFromSessionKeyRules	[]	[]
principalFromSessionKeyRules[].match	(unset)	(unset)
principalFromSessionKeyRules[].principal	(unset)	(unset)
namespacePolicies	[]	[]
namespacePolicies[].name	(unset)	(unset)
namespacePolicies[].readPrincipals	(unset)	(unset)
namespacePolicies[].writePrincipals	(unset)	(unset)
namespacePolicies[].includeInRecallByDefault	(unset)	(unset)
defaultRecallNamespaces	["self","shared"]	["self","shared"]
scopeProfiles	{}	unset unless running a hosted multi-user/project-aware deployment
defaultScopeProfile	(unset)	unset unless a scopeProfiles entry should control implicit reads/writes
teams	{}	trusted team membership for teamProject scope profile layers
cronRecallMode	all	all
cronRecallAllowlist	[]	[]
cronRecallPolicyEnabled	true	true
cronRecallNormalizedQueryMaxChars	480	480
cronRecallInstructionHeavyTokenCap	36	36
cronConversationRecallMode	auto	auto
autoPromoteToSharedEnabled	false	false
autoPromoteToSharedCategories	["fact","correction","decision","preference"]	["fact","correction","decision","preference"]
autoPromoteMinConfidenceTier	explicit	implied (recommended)
routingRulesEnabled	false	false
routingRulesStateFile	state/routing-rules.json	state/routing-rules.json
sharedContextEnabled	false	false unless you are actively using cross-agent memory sharing
sharedContextDir	(unset)	(unset)
sharedContextMaxInjectChars	4000	4000
crossSignalsSemanticEnabled	false	false
crossSignalsSemanticTimeoutMs	4000	4000
sharedCrossSignalSemanticEnabled	false	false
sharedCrossSignalSemanticTimeoutMs	4000	4000
sharedCrossSignalSemanticMaxCandidates	120	120
compoundingEnabled	false	false unless you are ready to curate weekly syntheses
compoundingWeeklyCronEnabled	false	false
compoundingSemanticEnabled	false	false
compoundingSynthesisTimeoutMs	15000	15000
compoundingInjectEnabled	true	true
factDeduplicationEnabled	true	true
semanticDedupEnabled	true	true (issue #373 — write-time semantic guard)
semanticDedupThreshold	0.92	0.92 (tighten to 0.95 for high-precision corpora, loosen to 0.88 for noisy transcripts)
semanticDedupCandidates	5	5
factArchivalEnabled	false	false unless you have validated archive policy on your corpus
factArchivalAgeDays	90	90
factArchivalMaxImportance	0.3	0.3
factArchivalMaxAccessCount	2	2
factArchivalProtectedCategories	["commitment","preference","decision","principle","procedure"]	["commitment","preference","decision","principle","procedure"]
lifecyclePolicyEnabled	false	false until you are ready to measure lifecycle outcomes
lifecycleFilterStaleEnabled	false	false for the initial lifecycle rollout
lifecyclePromoteHeatThreshold	0.55	0.55
lifecycleStaleDecayThreshold	0.65	0.65
lifecycleArchiveDecayThreshold	0.85	0.85
lifecycleProtectedCategories	["decision","principle","commitment","preference","procedure"]	["decision","principle","commitment","preference","procedure"]
lifecycleMetricsEnabled	false	true when lifecyclePolicyEnabled=true
procedural.enabled	true	true (default-on since issue #567 PR 4/5) or false to opt out. See procedural-memory.md.
procedural.minOccurrences	3	3 (use 0 only to intentionally disable mining; see narrative section)
procedural.successFloor	0.75	0.75 (raised from 0.7 in issue #567 PR 3/5)
procedural.autoPromoteOccurrences	8	8
procedural.autoPromoteEnabled	false	false until promotion rules are validated on your corpus
procedural.lookbackDays	14	14 (lowered from 30 in issue #567 PR 3/5)
procedural.proceduralMiningCronAutoRegister	false	false unless you intentionally want installer cron registration
procedural.recallMaxProcedures	2	2 (lowered from 3 in issue #567 PR 3/5)
patternReinforcementEnabled	false	false until you have enough cross-session data to observe clustering benefits. See pattern-reinforcement.md.
patternReinforcementCadenceMs	604800000	604800000 (7 days). Lower to 86400000 (1 day) for faster iteration during evaluation; set to 0 to disable cadence gating entirely.
patternReinforcementMinCount	3	3 (minimum meaningful pattern; clusters of 2 are allowed but 3 reduces false positives on small corpora)
patternReinforcementCategories	["preference", "fact", "decision"]	["preference", "fact", "decision"] (procedure excluded intentionally — procedural miner handles that category)
reinforcementRecallBoostEnabled	false	false until you confirm pattern reinforcement is producing high-quality canonicals. Enable recall boost only after observing remnic patterns list output.
reinforcementRecallBoostWeight	0.05	0.05 (per-unit score bonus per reinforcement_count; raise cautiously and pair with a lower reinforcementRecallBoostMax if you want fast saturation)
reinforcementRecallBoostMax	0.3	0.3 (a 30-point maximum additive boost; lower to 0.1–0.15 for conservative uplift)
proactiveExtractionEnabled	false	false until you validate the second pass in your environment
contextCompressionActionsEnabled	false	false unless you are validating action-policy flows
compressionGuidelineLearningEnabled	false	false unless action-policy telemetry is already stable
compressionGuidelineSemanticRefinementEnabled	false	false unless deterministic guideline learning is already stable
compressionGuidelineSemanticTimeoutMs	2500	2500
maxProactiveQuestionsPerExtraction	2	2
proactiveExtractionTimeoutMs	2500	2500
proactiveExtractionMaxTokens	900	900
proactiveExtractionCategoryAllowlist	(unset)	(unset)
maxCompressionTokensPerHour	1500	1500
behaviorLoopAutoTuneEnabled	false	false until you are ready for canary tuning
behaviorLoopLearningWindowDays	14	14
behaviorLoopMinSignalCount	10	10
behaviorLoopMaxDeltaPerCycle	0.1	0.1
behaviorLoopProtectedParams	["maxMemoryTokens","qmdMaxResults","qmdColdMaxResults","recallPlannerMaxQmdResultsMinimal","verbatimArtifactsMaxRecall"]	["maxMemoryTokens","qmdMaxResults","qmdColdMaxResults","recallPlannerMaxQmdResultsMinimal","verbatimArtifactsMaxRecall"]
searchBackend	qmd	qmd
remoteSearchBaseUrl	(unset)	(unset)
remoteSearchApiKey	(unset)	(unset)
remoteSearchTimeoutMs	30000	30000
lancedbEnabled	false	false
lanceDbPath	~/.openclaw/workspace/memory/local/lancedb	~/.openclaw/workspace/memory/local/lancedb
lanceEmbeddingDimension	1536	1536
meilisearchEnabled	false	false
meilisearchHost	http://localhost:7700	http://localhost:7700
meilisearchApiKey	(unset)	(unset)
meilisearchTimeoutMs	30000	30000
meilisearchAutoIndex	false	false
oramaEnabled	false	false
oramaDbPath	~/.openclaw/workspace/memory/local/orama	~/.openclaw/workspace/memory/local/orama
oramaEmbeddingDimension	1536	1536
qmdDaemonEnabled	true	true
qmdDaemonUrl	http://localhost:8181/mcp	http://localhost:8181/mcp
qmdDaemonRecheckIntervalMs	60000	60000
qmdIntentHintsEnabled	false	false
qmdExplainEnabled	false	false
knowledgeIndexEnabled	true	true
knowledgeIndexMaxEntities	40	40
knowledgeIndexMaxChars	4000	4000
entityRetrievalEnabled	true	true
entityRetrievalMaxChars	2400	2400
entityRetrievalMaxHints	2	2
entityRetrievalMaxSupportingFacts	6	6
entityRetrievalMaxRelatedEntities	3	3
entityRetrievalRecentTurns	6	6
entityRelationshipsEnabled	true	true
entityActivityLogEnabled	true	true
entityActivityLogMaxEntries	20	20
entityAliasesEnabled	true	true
entitySummaryEnabled	true	true
recallBudgetChars	8000	8000
recallPipeline	[{"id":"shared-context","enabled":false,"maxChars":4000},{"id":"profile","enabled":true,"consolidateTriggerLines":100,"consolidateTargetLines":50},{"id":"identity-continuity","enabled":false},{"id":"entity-retrieval","enabled":true,"maxChars":2400,"maxHints":2,"maxSupportingFacts":6,"maxRelatedEntities":3,"recentTurns":6},{"id":"knowledge-index","enabled":true,"maxChars":4000,"maxEntities":40},{"id":"verbatim-artifacts","enabled":false},{"id":"memory-boxes","enabled":false},{"id":"temporal-memory-tree","enabled":false},{"id":"objective-state","enabled":false,"maxResults":4,"maxChars":1800},{"id":"causal-trajectories","enabled":false,"maxResults":3,"maxChars":2200},{"id":"trust-zones","enabled":false,"maxResults":3,"maxChars":1800},{"id":"harmonic-retrieval","enabled":false,"maxResults":3,"maxChars":2200},{"id":"verified-episodes","enabled":false,"maxResults":3,"maxChars":1800},{"id":"verified-rules","enabled":false,"maxResults":3,"maxChars":1800},{"id":"work-products","enabled":false,"maxResults":3,"maxChars":1800},{"id":"memories","enabled":true,"maxResults":8},{"id":"compression-guidelines","enabled":false},{"id":"native-knowledge","enabled":false,"maxResults":4,"maxChars":2400},{"id":"transcript","enabled":true,"maxTurns":50,"maxTokens":1000,"lookbackHours":12},{"id":"summaries","enabled":true,"maxCount":6,"lookbackHours":24},{"id":"conversation-recall","enabled":false,"topK":3,"maxChars":2500,"timeoutMs":800},{"id":"compounding","enabled":false,"maxPatterns":40,"maxRubrics":4},{"id":"questions","enabled":false}]	[{"id":"shared-context","enabled":false,"maxChars":4000},{"id":"profile","enabled":true,"consolidateTriggerLines":100,"consolidateTargetLines":50},{"id":"identity-continuity","enabled":false},{"id":"entity-retrieval","enabled":true,"maxChars":2400,"maxHints":2,"maxSupportingFacts":6,"maxRelatedEntities":3,"recentTurns":6},{"id":"knowledge-index","enabled":true,"maxChars":4000,"maxEntities":40},{"id":"verbatim-artifacts","enabled":false},{"id":"memory-boxes","enabled":false},{"id":"temporal-memory-tree","enabled":false},{"id":"objective-state","enabled":false,"maxResults":4,"maxChars":1800},{"id":"causal-trajectories","enabled":false,"maxResults":3,"maxChars":2200},{"id":"trust-zones","enabled":false,"maxResults":3,"maxChars":1800},{"id":"harmonic-retrieval","enabled":false,"maxResults":3,"maxChars":2200},{"id":"verified-episodes","enabled":false,"maxResults":3,"maxChars":1800},{"id":"verified-rules","enabled":false,"maxResults":3,"maxChars":1800},{"id":"work-products","enabled":false,"maxResults":3,"maxChars":1800},{"id":"memories","enabled":true,"maxResults":8},{"id":"compression-guidelines","enabled":false},{"id":"native-knowledge","enabled":false,"maxResults":4,"maxChars":2400},{"id":"transcript","enabled":true,"maxTurns":50,"maxTokens":1000,"lookbackHours":12},{"id":"summaries","enabled":true,"maxCount":6,"lookbackHours":24},{"id":"conversation-recall","enabled":false,"topK":3,"maxChars":2500,"timeoutMs":800},{"id":"compounding","enabled":false,"maxPatterns":40,"maxRubrics":4},{"id":"questions","enabled":false}]
recallPipeline[].id	shared-context	shared-context
recallPipeline[].enabled	false	false
recallPipeline[].maxChars	4000	4000
recallPipeline[].consolidateTriggerLines	(unset)	(unset)
recallPipeline[].consolidateTargetLines	(unset)	(unset)
recallPipeline[].maxEntities	(unset)	(unset)
recallPipeline[].maxResults	(unset)	(unset)
recallPipeline[].maxTurns	(unset)	(unset)
recallPipeline[].maxTokens	(unset)	(unset)
recallPipeline[].lookbackHours	(unset)	(unset)
recallPipeline[].maxCount	(unset)	(unset)
recallPipeline[].topK	(unset)	(unset)
recallPipeline[].timeoutMs	(unset)	(unset)
recallPipeline[].maxPatterns	(unset)	(unset)
extractionJudgeEnabled	false	false
extractionJudgeModel	""	""
extractionJudgeBatchSize	20	20
extractionJudgeShadow	false	false
semanticChunkingEnabled	false	false
semanticChunkingConfig	{"targetTokens":200,"minTokens":100,"maxTokens":400,"smoothingWindowSize":3,"boundaryThresholdStdDevs":1.0,"embeddingBatchSize":32,"fallbackToRecursive":true}	{"targetTokens":200,"minTokens":100,"maxTokens":400,"smoothingWindowSize":3,"boundaryThresholdStdDevs":1.0,"embeddingBatchSize":32,"fallbackToRecursive":true}
semanticChunkingConfig.targetTokens	200	200
semanticChunkingConfig.minTokens	100	100
semanticChunkingConfig.maxTokens	400	400
semanticChunkingConfig.smoothingWindowSize	3	3
semanticChunkingConfig.boundaryThresholdStdDevs	1.0	1.0
semanticChunkingConfig.embeddingBatchSize	32	32
semanticChunkingConfig.fallbackToRecursive	true	true
versioningEnabled	false	false
versioningMaxPerPage	50	50
versioningSidecarDir	".versions"	".versions"
emitLegacyTools	true	true
citationsEnabled	false	false
citationsAutoDetect	true	true
taxonomyEnabled	false	false
taxonomyAutoGenResolver	true	true
enrichmentEnabled	false	false
enrichmentAutoOnCreate	false	false
enrichmentMaxCandidatesPerEntity	20	20
binaryLifecycleEnabled	false	false
binaryLifecycleGracePeriodDays	7	7
binaryLifecycleBackendType	"none"	"none"
binaryLifecycleBackendPath	""	""
codexMarketplaceEnabled	true	true
memoryExtensionsEnabled	true	true
memoryExtensionsRoot	""	""