Configuration reference

Lore’s configuration has three layers, in order of precedence (highest first):

1. Environment variables (the LORE_* family — see the env-vars reference page) — process-level overrides, used by the gateway CLI and runtime.
2. The .lore.json file in the project root — per-project, JSONC-parseable, the primary config surface for memory behavior, budget, distillation, search, knowledge, and cross-project settings. This page documents it.
3. Built-in defaults — every field is optional; the Zod schema supplies safe defaults for anything you omit.

Where .lore.json lives

The gateway looks for the .lore.json file in a single location:

• The project root — the directory the gateway was launched from, identified by the .lore.json file at its top level.

If no .lore.json is found, the gateway uses all defaults (every field is optional). In hosted mode (env var LORE_HOSTED_MODE=1), .lore.json is not read at all — a crafted file on a client-controlled path could alter gateway behavior, so the gateway uses the defaults to prevent tampering. Hosted-mode deployments should configure Lore through other means (admin-controlled env vars, fixed configuration baked into the deployment).

The file is JSONC (JSON with // line comments and slash-star block comments, plus trailing commas). Example:

{
  // Use a cheaper worker model for distillation
  "workerModel": { "providerID": "anthropic", "modelID": "claude-3-5-haiku-latest" },
  "curator": { "enabled": true, "maxEntries": 30 },
}

How to override a single field

The cleanest way to override a single field is via env var if Lore reads it, or by adding the field to .lore.json in the project root. See the env-vars reference page for the env-var override path; everything else is here in the .lore.json reference below.

Field reference

• model — Default session model. When omitted, Lore uses the model from the first client request.
• workerModel — Background-worker model for distillation, curation, and query expansion. Same-provider invariant: workers MUST use the same provider as the session.
• budget — Context-window budget fractions. Sum plus LTM ≈ 1.0.
• idleResumeMinutes — Minutes of inactivity after which Lore refreshes the byte-identity caches on resume (upstream prompt cache is cold). 5 = matches Anthropic’s default-tier TTL. Set to 60 for extended (1h) cache tier. 0 to disable. Default: 5.
• timeouts — Foreground and background provider request deadlines, in milliseconds. Non-hosted gateways read these values from .lore.json; hosted gateways use environment variables.
• distillation — Distillation pipeline tuning (segment size, thresholds, tool-output truncation).
• knowledge — Long-term knowledge (curator, entity injection) controls.
• curator — Curator scheduling and consolidation thresholds.
• pruning — Storage retention and emergency-pruning thresholds.
• search — Recall and search pipeline tuning: FTS weights, query expansion, vector boost, embeddings, and output formatting.
• cache — Anthropic prompt cache TTL and speculative warming.
• workspaces — Workspace sub-project paths or globs (relative to .lore.json). Imported into the root knowledge base on startup. Supports literal paths and single-level globs (e.g. “packages/*”).
• crossProject — Include cross-project knowledge in compaction summaries and auto-promote knowledge that recurs across 3+ projects. Default: true.
• agentsFile — AGENTS.md/CLAUDE.md export/import configuration.
• loreFile — .lore.md export/import configuration.
• user — User identity for the self-entity. Falls back to git config user.name / user.email if omitted.
• invariantCheck — lore lint (semantic linter) tuning.

model

Default session model. When omitted, Lore uses the model from the first client request.

Field	Type	Default	Constraints	Description
providerID	string	—		Provider ID (e.g. ‘anthropic’, ‘openai’).
modelID	string	—		Model identifier within the provider.

workerModel

Background-worker model for distillation, curation, and query expansion. Same-provider invariant: workers MUST use the same provider as the session.

Field	Type	Default	Constraints	Description
providerID	string	—		Provider ID for the worker model.
modelID	string	—		Model identifier within the provider.

budget

Context-window budget fractions. Sum plus LTM ≈ 1.0.

Field	Type	Default	Constraints	Description
distilled	number	0.25	min 0.05, max 0.5	Fraction of usable context for distilled prefix. Default: 0.25.
raw	number	0.4	min 0.1, max 0.7	Fraction of usable context for the recent raw window (un-distilled). Default: 0.4.
output	number	0.25	min 0.1, max 0.5	Fraction of usable context reserved for model output. Default: 0.25.
ltm	number	0.05	min 0.02, max 0.3	Max fraction of usable context reserved for context-bound LTM system-prompt injection. Default: 0.05 (5%).
preferenceLtm	number	0.02	min 0.01, max 0.1	Fraction of usable context for stable LTM (preferences). Independent of ltm. Default: 0.02 (2%).
targetCacheReadCostPerTurn	number	0.1	min 0	Per-turn cache-read cost target in dollars. Controls when layer 0 escalates to layer 1. Default: 0.10. Set to 0 to disable cost-aware capping.
maxLayer0Tokens	number	—	min 0	Direct override for the layer-0 token cap. 0 = disabled (use full context). Default: undefined (use cost-aware auto).
targetBustCost	number	1	min 0	@deprecated Ignored. Tier-based bust-vs-continue replaces static cap.
maxContextTokens	number	—	min 0	@deprecated Ignored. Tier-based bust-vs-continue replaces static cap.
qualityKnee	number	—	min 0, max 1	Manual override for the per-model quality knee (fill fraction where compression ramps). (0,1). Default: undefined (use the built-in per-model-family seed table, fallback 0.4).

idleResumeMinutes

Minutes of inactivity after which Lore refreshes the byte-identity caches on resume (upstream prompt cache is cold). 5 = matches Anthropic’s default-tier TTL. Set to 60 for extended (1h) cache tier. 0 to disable. Default: 5.

timeouts

Foreground and background provider request deadlines, in milliseconds. Non-hosted gateways read these values from .lore.json; hosted gateways use environment variables.

Field	Type	Default	Constraints	Description
foregroundSseInactivityMs	number	—	min 1000, max 2147423647	Foreground upstream SSE inactivity deadline in milliseconds. Capped at 2147423647ms to reserve 60000ms of request-timeout headroom. Default: 600000. Environment variable LORE_FOREGROUND_SSE_INACTIVITY_MS takes precedence.
foregroundRequestTimeoutMs	number	—	min 1000, max 2147483647	Foreground whole-request deadline in milliseconds. Default: 900000. Automatically raised to at least 60000ms above foregroundSseInactivityMs. Environment variable LORE_FOREGROUND_REQUEST_TIMEOUT_MS takes precedence.
memoryPreparationTimeoutMs	number	—	min 1000, max 60000	Post-decode memory preparation budget in milliseconds. Default: 8000. LORE_MEMORY_PREPARATION_TIMEOUT_MS takes precedence at request admission. A verified accepted source window and knowledge pin may continue on optional refresh expiry; otherwise preparation returns a retryable error.
workerResponseInactivityMs	number	—	min 1000, max 2147423647	Background worker upstream response inactivity deadline in milliseconds. Capped at 2147423647ms to reserve 60000ms of request-timeout headroom. Default: 600000. Environment variable LORE_WORKER_RESPONSE_INACTIVITY_MS takes precedence.
workerRequestTimeoutMs	number	—	min 1000, max 2147483647	Background worker whole-request deadline in milliseconds. Default: 900000. Automatically raised to at least 60000ms above workerResponseInactivityMs. Environment variable LORE_WORKER_REQUEST_TIMEOUT_MS takes precedence.

distillation

Distillation pipeline tuning (segment size, thresholds, tool-output truncation).

Field	Type	Default	Constraints	Description
minMessages	number	5	min 3	Minimum number of messages before a segment is eligible for distillation. Default: 5.
minSegmentTokens	number	64	min 16	Minimum tokens for a segment to be worth distilling. Default: 64.
maxSegmentTokens	number	16384	min 256	Maximum tokens per distillation segment before splitting. Default: 16384.
metaThreshold	number	20	min 3	Number of gen-0 segments that triggers meta-distillation. Default: 20.
toolOutputMaxChars	number	4000	min 0	Max chars per tool output for distillation input. Set to 0 to disable truncation. Default: 4000.
userBlobMaxChars	number	12000	min 0	Trigger threshold (chars) for embedding-based user-blob reduction. Set to 0 to disable. Default: 12000.
userBlobKeepChars	number	6000	min 0	Chars kept after reducing an oversized user blob. Default: 6000.
userBlobMaxSegments	number	48	min 1	Max blob segments embedded during user-blob reduction. Default: 48.
userBlobHeadChars	number	1500	min 0	Chars of the leading user-prose prefix always kept during blob reduction. Set to 0 to disable. Default: 1500.
recentSegmentsToKeep	number	5	min 0	Number of most-recent gen-0 segments to keep un-archived when meta-distillation fires. Default: 5.

knowledge

Long-term knowledge (curator, entity injection) controls.

Field	Type	Default	Constraints	Description
enabled	boolean	true		Enable long-term knowledge storage and system-prompt injection. When false, the curator is disabled but recall and context management remain active. Default: true.
maxEntityInject	number	30	min 0	Max entities to inject into the agent system prompt. Set to 0 to disable. Default: 30.
autoToolFailureGotchas	boolean	false		Auto-create gotcha entries from recurring tool failures. Default false (noisy; can churn the LTM cache).
outcomeReward	boolean	true		Adjust knowledge confidence by within-session verifier (test/build/typecheck/lint) outcomes. Default: true.
referenceValidation	boolean	true		Lower confidence on entries whose file:line / command references no longer resolve against the repo. Unverifiable refs never penalize. Default: true.
contextSources	array	["distillation"]		Fold relevance-ranked distillation/temporal memory into the context-bound injection so facts are passively present (no recall tool needed). Default: [“distillation”]; add “temporal” for raw messages; [] = off.
minRelevance	number	0.35	min 0, max 1	Minimum cosine similarity for a vector-only knowledge match to be surfaced into a session. FTS keyword matches bypass it. 0 disables. Default: 0.35.
architectureInjection	enum	"ranked"		Architecture in passive knowledge injection: “ranked” (default) competes by relevance with other entries; “reserved” restores the 20% first-pass allocation. Explicit category reads and recall remain available.
taskFilter	boolean	true		Filter contextual knowledge with checkpoint-time task anchors after budget packing. Keeps preferences and explicit recall available. Set false to retain implicit semantic-only matches and contextual decisions/patterns. Default: true.

curator

Curator scheduling and consolidation thresholds.

Field	Type	Default	Constraints	Description
enabled	boolean	true		Enable the curator (knowledge extraction from conversation). Default: true.
onIdle	boolean	true		Run the curator on session idle (in addition to turn-based). Default: true.
inFlight	boolean	false		Run the curator mid-conversation (turn-based), not just on idle. Default: false. WARNING: only enable on free-write / non-caching providers (e.g. MiniMax). On cache-sensitive providers (Anthropic), mid-session curation changes the knowledge base, which queues a knowledge-delta block (user→assistant pair appended mid-conversation) that re-writes part of the prompt tail each time the knowledge set changes. Deferring curation to idle lets the queue stay empty during active turns. Where cache writes are free this is harmless and yields fresher knowledge sooner.
afterTurns	number	3	min 1	Minimum turns between curator runs. Default: 3.
maxEntries	number	200	min 10	Per-project knowledge entry ceiling. A generous backstop, not a quality gate: injection is already token-budget-capped, and the confidence lifecycle (decay + reinforcement) governs what stays. When exceeded, the lowest-value entries are evicted. Default: 200.
contextTokenBudget	number	20000	min 2000	Token budget for the existing-entries context sent to the curator each run. Bounds curator LLM cost so it stops scaling with stored entry count; a generous safety ceiling that only trims pathological sets (cross-project entries are always kept). Default: 20000.

pruning

Storage retention and emergency-pruning thresholds.

Field	Type	Default	Constraints	Description
retention	number	120	min 1	Days to keep distilled temporal messages before pruning. Default: 120.
maxStorage	number	1024	min 50	Max total temporal_messages storage in MB before emergency pruning. Default: 1024 (1 GB).

search

Recall and search pipeline tuning: FTS weights, query expansion, vector boost, embeddings, and output formatting.

Field	Type	Default	Constraints	Description
ftsWeights	object	{"title":6,"content":2,"category":3}		BM25 column weights for knowledge FTS5 [title, content, category].
recallLimit	number	10	min 1, max 50	Max results per source in recall tool before fusion. Default: 10.
queryExpansion	boolean	true		Enable LLM-based query expansion (2-3 alternative phrasings) for the recall tool. Guarded by a 3s timeout. Default: true.
queryExpansionMaxTerms	number	8	min 2, max 20	Max query terms (after stopword removal) for LLM expansion. Longer queries skip expansion. Default: 8.
vectorBoostWeight	number	1.5	min 1, max 5	RRF weight multiplier for vector search lists (when query has enough terms). Set to 1.0 to disable. Default: 1.5.
vectorBoostMinTerms	number	2	min 1, max 10	Minimum meaningful query terms (after stopword removal) to activate vector boost. Default: 2.
graphExpansion	boolean	true		Enable entity-graph fan-in (linked knowledge + 1-hop relation neighbors) for the recall tool. Default: true.
fuzzyRecall	boolean	true		Append a low-weight fuzzy title leg to recall when FTS + vector underfill recallLimit. Default: true.
graphBoostWeight	number	1	min 0, max 5	RRF weight multiplier for entity-graph fan-in lists. Set to 0 to neutralize. Default: 1.0.
embeddings	object	{"enabled":true,"provider":"local","model":"nomic-ai/nomic-embed-text-v1.5","dimensions":768,"queryTimeoutMs":1500,"workerOffload":true,"workerPoolSize":2}		Vector embedding search provider, model, and dimensions.
recall	object	{"charBudget":12000,"relevanceFloor":0.15,"maxResults":15,"absoluteFloor":0,"chainMaxExecutions":24}		Recall output formatting and result-count limits.

search.ftsWeights

BM25 column weights for knowledge FTS5 [title, content, category].

Field	Type	Default	Constraints	Description
title	number	6	min 0	BM25 weight for the entry title column. Default: 6.0.
content	number	2	min 0	BM25 weight for the entry content column. Default: 2.0.
category	number	3	min 0	BM25 weight for the entry category column. Default: 3.0.

search.embeddings

Vector embedding search provider, model, and dimensions.

Field	Type	Default	Constraints	Description
enabled	boolean	true		Enable vector embedding search. Set to false to explicitly disable. Default: true.
provider	enum	"local"		Embedding provider. “local” (no API key, on-device), “voyage” (VOYAGE_API_KEY), “openai” (OPENAI_API_KEY). Default: “local”.
model	string	"nomic-ai/nomic-embed-text-v1.5"		Model ID for the embedding provider. Default depends on provider.
dimensions	number	768	min 64, max 2048	Embedding dimensions. Default: 768 (local) / 1024 (voyage) / 1536 (openai). Local Nomic v1.5 supports Matryoshka: 64, 128, 256, 512, 768.
queryTimeoutMs	number	1500	min 100, max 60000	Maximum foreground LTM query-embedding wait in milliseconds, including worker startup and queueing. FTS is used on expiry. Default: 1500.
workerOffload	boolean	true		Run heavy reads on a read-worker pool off the main event loop. Default true; when disabled, required memory preparation returns a retryable 503 rather than scanning on the gateway thread.
workerPoolSize	number	2	min 1, max 16	Number of read-worker threads for off-thread vector search. Default: 2.
embedPoolSize	number	—	min 1, max 8	Number of local embedding worker threads (each loads its own ~137MB model). Default: memory-gated (1–2). Override with LORE_EMBED_POOL_SIZE.
backfillCpuDuty	number	—	min 0.1, max 1	Legacy temporal backfill duty setting. Durable temporal admission no longer performs inference; the bounded scheduler controls embedding concurrency.

search.recall

Recall output formatting and result-count limits.

Field	Type	Default	Constraints	Description
charBudget	number	12000	min 2000, max 20000	Total character budget for recall output (~3K tokens at 12000 chars). Default: 12000.
relevanceFloor	number	0.15	min 0, max 1	Minimum RRF score (relative to top) to keep. Set to 0 to disable. Default: 0.15.
maxResults	number	15	min 3, max 30	Max results to show in recall output. Default: 15.
absoluteFloor	number	0	min 0	Absolute RRF score floor; drops weak matches even via the keep-3 backfill. Default: 0 (disabled).
chainMaxExecutions	number	24	min 12, max 128	Emergency execution cap for one recall chain. Default: 24; normal resource and stall budgets stop first.

cache

Anthropic prompt cache TTL and speculative warming.

Field	Type	Default	Constraints	Description
conversationTTL	enum	"auto"		Conversation cache breakpoint TTL. “5m” (Anthropic standard), “1h” (extended tier, 2× write cost), “auto” (auto-upgrade based on cold-cache rate). Default: “auto”.
warming	object	{"enabled":true}		Speculative cache warming (keepalive requests before TTL expiry).

cache.warming

Speculative cache warming (keepalive requests before TTL expiry).

Field	Type	Default	Constraints	Description
enabled	boolean	true		Enable cache warming. Default: true.
minReturnProbability	number	—	min 0, max 1	Override return probability threshold below which warming is skipped. Default: auto-derived from cost ratio.

workspaces

Workspace sub-project paths or globs (relative to .lore.json). Imported into the root knowledge base on startup. Supports literal paths and single-level globs (e.g. “packages/*”).

crossProject

Include cross-project knowledge in compaction summaries and auto-promote knowledge that recurs across 3+ projects. Default: true.

agentsFile

AGENTS.md/CLAUDE.md export/import configuration.

Field	Type	Default	Constraints	Description
enabled	boolean	true		Enable AGENTS.md export/import behaviour. Set to false to disable. Default: true.
path	string	"auto"		Path to the agents file, relative to the project root, or ‘auto’ (default) to write CLAUDE.md for Claude Code sessions and AGENTS.md otherwise. Set an explicit path (e.g. ‘AGENTS.md’ or ‘CLAUDE.md’) to override.

loreFile

.lore.md export/import configuration.

Field	Type	Default	Constraints	Description
enabled	boolean	true		Set to false to disable .lore.md export/import. When disabled, .lore.md is not written, startup skips the .lore.md import branch, the recall tool omits the .lore.md commit reminder, and the file watcher ignores .lore.md. Knowledge stays in the database and is still injected into the system prompt via LTM. Pair with agentsFile.enabled=true to keep sharing knowledge via an inline section in AGENTS.md. Default: true.

user

User identity for the self-entity. Falls back to git config user.name / user.email if omitted.

Field	Type	Default	Constraints	Description
name	string	—		Display name. Overrides git config user.name.
email	string	—		Email address. Overrides git config user.email.
aliases	array<{ type, value }>	[]		Additional aliases for the self entity.
metadata	Record<string, unknown>	—		Metadata for the self entity (description, role, notes, etc.).

invariantCheck

lore lint (semantic linter) tuning.

Field	Type	Default	Constraints	Description
effort	enum	"off"		Reasoning effort for the lint judge (off|low|medium|high|xhigh). Trades cost for depth on reasoning-capable models. Default: off. Override per-run with –effort.