memory.json.example

{
  "_comment": "Quaid Memory Plugin — Configuration Template. Copy to config/memory.json and customize.",
  "adapter": {
    "type": "standalone",
    "_notes": "Required. Set to 'standalone' for local CLI installs, 'openclaw' when running under OpenClaw, or the relevant adapter id for other host integrations."
  },

  "plugins": {
    "enabled": true,
    "strict": true,
    "apiVersion": 1,
    "paths": ["plugins"],
    "allowList": [],
    "slots": {
      "adapter": "",
      "ingest": [],
      "dataStores": []
    },
    "_notes": "Phase-1 plugin runtime foundation. paths are scanned for plugin.json manifests. allowList empty means all discovered plugins are eligible. slots define active plugin IDs by zone. strict=true fails on invalid manifests/conflicts."
  },

  "systems": {
    "memory": true,
    "journal": true,
    "projects": true,
    "workspace": true,
    "_notes": "Toggle subsystems on/off. memory: fact extraction+recall, journal: personality snippets+reflective log, projects: auto-update project docs, workspace: core markdown health monitoring. Infrastructure tasks (backup, tests, cleanup) always run."
  },

  "models": {
    "llmProvider": "default",
    "deepReasoningProvider": "default",
    "fastReasoningProvider": "default",
    "embeddingsProvider": "ollama",
    "fastReasoning": "default",
    "deepReasoning": "default",
    "deepReasoningModelClasses": {
      "openai-codex": "gpt-5.3-codex",
      "openai": "gpt-5.3-codex",
      "anthropic": "claude-opus-4-6"
    },
    "fastReasoningModelClasses": {
      "openai-codex": "gpt-5.1-codex-mini",
      "openai": "gpt-5.1-codex-mini",
      "anthropic": "claude-haiku-4-5"
    },
    "fastReasoningContext": 200000,
    "deepReasoningContext": 200000,
    "fastReasoningMaxOutput": 8192,
    "deepReasoningMaxOutput": 16384,
    "batchBudgetPercent": 0.5,
    "_notes": "llmProvider picks the shared default provider. deepReasoningProvider/fastReasoningProvider can override per tier. deepReasoning/fastReasoning allow explicit model IDs or 'default' (mapped via *ReasoningModelClasses for the effective provider)."
  },

  "capture": {
    "enabled": true,
    "strictness": "high",
    "inactivityTimeoutMinutes": 120,
    "autoCompactionOnTimeout": true,
    "skipPatterns": [
      "^(thanks|ok|sure|yes|no)$",
      "^(hi|hello|hey)\\b"
    ],
    "_notes": "strictness: high = only clear facts, medium = default, low = capture more. inactivityTimeoutMinutes: extract after inactivity. autoCompactionOnTimeout: after timeout extraction, compact session to reduce token usage. skipPatterns: regex patterns to skip during extraction."
  },

  "decay": {
    "enabled": true,
    "thresholdDays": 30,
    "ratePercent": 10,
    "minimumConfidence": 0.1,
    "protectVerified": true,
    "protectPinned": true,
    "reviewQueueEnabled": true,
    "mode": "exponential",
    "baseHalfLifeDays": 60,
    "accessBonusFactor": 0.15,
    "_notes": "Ebbinghaus exponential decay. baseHalfLifeDays: half-life for facts with 0 accesses. Each access extends half-life by accessBonusFactor (15%). thresholdDays: minimum age before decay starts."
  },

  "janitor": {
    "enabled": true,
    "dryRun": false,
    "applyMode": "auto",
    "parallel": {
      "enabled": true,
      "llmWorkers": 4,
      "llmWorkersByTask": {
        "review_pending": 2,
        "dedup_review": 2,
        "decay_review": 2,
        "contradiction_resolution": 2
      },
      "lifecyclePrepassEnabled": true,
      "lifecyclePrepassWorkers": 3,
      "lockEnforcementEnabled": true,
      "lockWaitSeconds": 120,
      "lockRequireRegistration": true
    },
    "approvalPolicies": {
      "coreMarkdownWrites": "ask",
      "projectDocsWrites": "ask",
      "destructiveMemoryOps": "auto"
    },
    "taskTimeoutMinutes": 60,
    "opusReview": {
      "enabled": true,
      "batchSize": 50,
      "maxTokens": 4000
    },
    "dedup": {
      "similarityThreshold": 0.85,
      "highSimilarityThreshold": 0.95,
      "autoRejectThreshold": 0.98,
      "grayZoneLow": 0.88,
      "haikuVerifyEnabled": true
    },
    "contradiction": {
      "enabled": true,
      "timeoutMinutes": 60,
      "minSimilarity": 0.6,
      "maxSimilarity": 0.85
    },
    "_notes": "Nightly maintenance pipeline. parallel is the global janitor parallelism control (on by default): llmWorkers controls datastore LLM batch concurrency and can be overridden per task via llmWorkersByTask; lifecyclePrepass controls janitor dry-run prepass concurrency. lockEnforcementEnabled requires registered write locks before maintenance routines can modify flat files/DB; lockRequireRegistration blocks routines that do not declare write resources; lockWaitSeconds controls lock timeout. approvalPolicies lets you set ask/auto per sensitive scope. Root markdown writes are usually ask; projects/quaid markdown is auto-managed by janitor. applyMode is a legacy master fallback (auto/ask/dry_run). opusReview: LLM reviews pending facts. dedup: similarity thresholds for duplicate detection. contradiction: detect conflicting facts."
  },

  "retrieval": {
    "defaultLimit": 5,
    "maxLimit": 8,
    "minSimilarity": 0.6,
    "notifyMinSimilarity": 0.85,
    "boostRecent": true,
    "boostFrequent": true,
    "maxTokens": 2000,
    "reranker": {
      "enabled": true,
      "provider": "llm",
      "model": "qwen2.5:7b",
      "topK": 20,
      "instruction": "Given a personal memory query, determine if this memory is relevant to the query",
      "_notes": "provider: 'llm' uses the configured fast-reasoning model (~$0.002/recall). 'ollama' uses the local model specified in 'model' field."
    },
    "rrfK": 60,
    "rerankerBlend": 0.5,
    "compositeRelevanceWeight": 0.60,
    "compositeRecencyWeight": 0.20,
    "compositeFrequencyWeight": 0.15,
    "multiPassGate": 0.70,
    "mmrLambda": 0.7,
    "coSessionDecay": 0.6,
    "recencyDecayDays": 90,
    "preInjectionPass": true,
    "routerFailOpen": true,
    "failHard": true,
    "traversal": {
      "useBeam": true,
      "beamWidth": 5,
      "maxDepth": 2,
      "hopDecay": 0.7,
      "_notes": "BEAM search for graph traversal. hopDecay: score multiplier per hop depth."
    },
    "_notes_router": "routerFailOpen=true (recommended default): if total_recall prepass cannot produce valid structured JSON after retry, log loudly and continue with deterministic default recall plan. Set false for strict fail-through behavior. If failHard=true, fallback behavior should be disabled.",
    "_notes_fail_hard": "failHard=true (recommended in dev/benchmark): disable fallbacks and fail fast on provider/path errors. Set false only when you explicitly want noisy degrade behavior."
  },

  "logging": {
    "enabled": true,
    "level": "info",
    "retentionDays": 30,
    "components": ["memory", "janitor"]
  },

  "notifications": {
    "level": "normal",
    "janitor": { "verbosity": "summary", "channel": "last_used" },
    "extraction": { "verbosity": "summary", "channel": "last_used" },
    "retrieval": { "verbosity": "off", "channel": "last_used" },
    "fullText": true,
    "showProcessingStart": true,
    "_notes_level": "Master verbosity: 'quiet' (errors only), 'normal' (summaries), 'verbose' (more detail), 'debug' (everything). Default: normal.",
    "_notes_features": "Per-feature config — verbosity: 'off', 'summary', or 'full' (null = inherit master level). channel: 'last_used' (follow user's active channel) or a specific channel name like 'telegram', 'discord', 'slack'. Shorthand: set to just a string like 'summary' for verbosity-only (no channel override).",
    "_notes_defaults": "quiet: all off. normal: janitor=summary, extraction=off, retrieval=summary. verbose: all summary. debug: all full.",
    "_notes_tip": "You can ask your agent to adjust verbosity or channel routing at any time, e.g. 'send janitor reports to discord' or 'make memory notifications quieter'."
  },

  "docs": {
    "autoUpdateOnCompact": true,
    "maxDocsPerUpdate": 3,
    "stalenessCheckEnabled": true,
    "updateTimeoutSeconds": 120,
    "coreMarkdown": {
      "enabled": true,
      "monitorForBloat": true,
      "monitorForOutdated": true,
      "files": {
        "SOUL.md": {
          "purpose": "Personality, vibe, values, interaction style",
          "maxLines": 80
        },
        "USER.md": {
          "purpose": "About the user — who they are, preferences, context",
          "maxLines": 150
        },
        "MEMORY.md": {
          "purpose": "Core memories loaded every session",
          "maxLines": 100
        },
        "TOOLS.md": {
          "purpose": "API docs, tool definitions, system configs",
          "maxLines": 150
        }
      },
      "_notes": "Janitor monitors these files for bloat and outdated info. These load every turn — keep them concise."
    },
    "journal": {
      "enabled": true,
      "snippetsEnabled": true,
      "mode": "distilled",
      "injectFull": false,
      "journalDir": "journal",
      "targetFiles": ["SOUL.md", "USER.md", "MEMORY.md"],
      "maxEntriesPerFile": 50,
      "maxTokens": 8192,
      "distillationIntervalDays": 7,
      "archiveAfterDistillation": true,
      "_notes": "Dual system: snippets (fast path, bullet points merged into core markdown nightly) + journal (slow path, reflective paragraphs distilled weekly). mode: 'distilled' = journal not loaded into context, 'full' = journal injected every turn. injectFull: EXPERIMENTAL — inject full journal text into agent context every turn (warning: journal size is uncapped, may consume significant context window)."
    },
    "sourceMapping": {},
    "docPurposes": {},
    "_notes": "sourceMapping: maps source files to their documentation files. docPurposes: describes what each doc covers. Both populated automatically by project discovery."
  },

  "projects": {
    "enabled": true,
    "projectsDir": "projects/",
    "stagingDir": "projects/staging/",
    "definitions": {},
    "defaultProject": "default",
    "_notes": "Projects are auto-discovered from the projects/ directory. You can also define them manually here."
  },

  "users": {
    "defaultOwner": "default",
    "identities": {
      "default": {
        "channels": {
          "cli": ["*"]
        },
        "speakers": ["User"],
        "personNodeName": "User"
      }
    },
    "_notes": "The setup script populates this with your actual identity. personNodeName is used for the Person node in the memory graph."
  },

  "database": {
    "path": "data/memory.db",
    "archivePath": "data/memory_archive.db",
    "walMode": true
  },

  "ollama": {
    "url": "http://localhost:11434",
    "embeddingModel": "nomic-embed-text",
    "embeddingDim": 768,
    "_WARNING_BEFORE_CHANGING_MODEL": [
      "STOP! Changing embeddingModel invalidates ALL existing embeddings.",
      "You MUST re-embed everything after changing: run 'quaid re-embed'.",
      "embeddingDim must match the new model's output dimension.",
      "Common models: nomic-embed-text (768, ~1.5GB RAM), qwen3-embedding:8b (4096, ~6GB), all-minilm (384, ~500MB)."
    ]
  },

  "rag": {
    "docsDir": "docs",
    "chunkMaxTokens": 800,
    "chunkOverlapTokens": 100,
    "maxResults": 5,
    "searchLimit": 5,
    "minSimilarity": 0.3
  }
}