config.example.yaml

vault:
  path: ~/Documents/Obsidian/MyVault
  # `Agent` is the folder the writeback jobs own; exclude it from ingest so
  # the daemon doesn't read its own memory files back in.
  exclude_dirs: [".obsidian", ".trash", "templates", "Agent"]
  # Frontmatter keys consulted for a note's episodic date, in order.
  date_frontmatter_keys: ["occurred_at", "date", "created", "created_at"]
  # Optional. Set to your vault's earliest bulk-import date to let otherwise
  # undated notes take their date from mtime. Run
  # `daemon migrate backfill-dates --dry-run` to see the detected import dates.
  # mtime_trusted_before: 2026-06-21

chunking:
  max_tokens: 512
  overlap_tokens: 50
  split_on_headers: ["#", "##", "###"]

embeddings:
  model: BAAI/bge-small-en-v1.5
  batch_size: 32
  device: auto  # auto | cpu | cuda | mps
  cache_folder: ./data/models   # model files live here; first run downloads, later runs are offline
  # Hybrid retrieval: fuse a BM25-style sparse signal with the dense one server-side in Qdrant.
  # Flip to false to A/B against dense-only. Switching modes requires `daemon ingest --full`.
  hybrid: false
  sparse_model: Qdrant/bm42-all-minilm-l6-v2-attentions

vector_store:
  backend: qdrant
  qdrant:
    # Embedded mode (the default): qdrant runs *inside* the daemon process
    # against this folder. No Docker, no server to start — just a directory
    # that persists across restarts. Back it up with the rest of ./data/.
    # (Deliberately *not* ./data/qdrant — that folder is the docker server's
    # storage mount, and the two on-disk layouts are not interchangeable.)
    path: ./data/qdrant-local
    # Embedded mode is single-process: exactly one daemon may hold the folder
    # at a time, so the CLI and the GUI can't both be open against it. Switch
    # to the server below if you hit either wall:
    #   * concurrent access (GUI + CLI + Hermes at once, or a shared machine)
    #   * a large vault — the server has real payload indexes and a much
    #     faster filtered search; embedded filters in Python.
    # To switch: comment out `path` above, uncomment `url` below, and run
    # `docker-compose up -d qdrant`. Setting both is a config error, on
    # purpose — silently preferring one would hide which store your vectors
    # actually landed in. Your existing index does not move between modes;
    # re-run `daemon ingest --full` after switching.
    # url: http://localhost:6333
    collection: chunks

graph:
  path: ./data/graph.gpickle
  manifest_path: ./data/manifest.json
  expansion_depth: 2
  distance_decay: 0.5

retrieval:
  seed_top_k: 8
  context_token_budget: 6000
  candidate_pool: 3   # always surface at least N candidates to the user
  # Build the candidate pool by team draft between the seed ranking and the
  # graph-expansion ranking, instead of by raw score. An expanded chunk scores
  # its seed's score decayed by graph distance, so under a score sort expansion
  # can never outrank its own seed and the expansion policy is never examined —
  # which is why picks used to teach the graph nothing. `daemon policy` shows
  # which ranking your picks actually favour. Set false for the old score order.
  interleave: true
  # Rerank the candidate pool with a local cross-encoder and enter that
  # ordering as a third drafting team (§IV.18), so `daemon policy` can report
  # whether reranking actually earns picks on this vault. Off by default —
  # enabling it without the weights cached would download on first query — so
  # run `daemon models download` before flipping this on.
  # rerank: true
  # rerank_model: cross-encoder/ms-marco-MiniLM-L-6-v2
  # rerank_max_candidates: 100

llm:
  provider: anthropic
  # Sonnet 4.6 is the cost/quality sweet spot for personal RAG synthesis.
  # Alternatives: claude-opus-4-7 (highest fidelity, ~5x cost), claude-haiku-4-5 (fastest, lowest cost).
  model: claude-sonnet-4-6
  max_tokens: 2048
  # Opus 4.7 and other reasoning-capable models reject `temperature`. Leave it
  # unset (null) for those; set a float (e.g. 0.3) for Sonnet/Haiku if desired.
  temperature: null
  # Background-agent jobs (extract / link / reflect) use this cheaper model.
  # Leave null to reuse `model` above; default Haiku 4.5 is ~15x cheaper than Opus.
  batch_model: claude-haiku-4-5

agent:
  # Master switch. Start in dry-run mode (daemon link --dry-run, daemon extract --note <one>)
  # to verify thresholds against your vault, then flip to true.
  enabled: true
  folder_name: Agent
  # Linker thresholds. The auto-apply gate is intentionally strict: cosine >= 0.85
  # AND the target's title must appear verbatim (case-insensitive) in the source body.
  # Everything between suggest and apply thresholds lands in Agent/link-suggestions-*.md.
  link_apply_cosine: 0.85
  link_apply_requires_title_substring: true
  link_suggest_cosine: 0.78
  link_max_per_note: 5
  tag_apply_min_neighbor_count: 4
  # Extractor.
  extract_min_word_count: 80
  extract_skip_if_processed_within_hours: 18
  # Reflection: which themes to maintain in Agent/memory-*.md.
  reflect_themes: [personality, projects, relationships, themes]
  reflect_lookback_days: 7
  # Don't touch a note that was saved within this many minutes (avoid collisions).
  write_grace_minutes: 30

  # --- Observer (M4) -----------------------------------------------------
  # The observer reads structural reports from M3 and writes a markdown
  # letter into <vault>/Agent/observer-<date>.md. Gated separately from
  # `enabled` so you can opt in after watching reflect/extract/link first.
  observer_enabled: false
  observer_lookback_days: 7
  observer_max_communities: 8
  # Opus 4.7 is the recommended model for the observer — interpreting
  # structural patterns into prose is the highest-quality lift in this
  # project. Set null to fall back to `llm.batch_model`.
  observer_model: claude-opus-4-7
  # How many prior letters to feed in for continuity, and how many to
  # surface in the rolling `observer.md` index.
  observer_prior_letters: 4
  observer_index_window: 8
  # Retention for snapshots auto-created by `daemon consolidate`.
  observer_snapshot_retention_days: 14

hermes:
  # My Daemon as the memory & dream layer for the Hermes agent (PLAN-HERMES.md).
  # Everything defaults off — the provider only activates when Hermes loads it
  # AND provider_enabled is true, and nothing is written unless allow_write_back.

  # --- Memory-provider plugin (Path A, primary) ---
  provider_enabled: false
  # Conversation captures land here: a *normal* (ingest-visible) folder, NOT the
  # daemon-owned Agent/, so captures become recallable in later sessions.
  capture_folder: Conversations/hermes
  capture_requires_confirmation: true
  # Skip capturing turns shorter than this (combined user+assistant chars).
  capture_min_chars: 120
  prefetch_budget_chars: 4000
  dream_block_budget_chars: 3000
  # One-paragraph identity framing injected above last night's letter at session
  # start. Leave blank for the built-in default.
  identity: ""
  # --- Shared retrieval defaults handed to the agent ---
  # Recall is retrieval-only by default: Hermes has its own model and wants
  # grounded, cited context, not an answer composed for it.
  recall_synthesize_default: false
  recall_top_k: 8
  # Master write-back gate (endorse + remember). Read-only until you flip this.
  allow_write_back: false
  # --- MCP server (Path B, optional portability) ---
  mcp_enabled: false
  mcp_transport: stdio   # stdio | http
  mcp_host: 127.0.0.1
  mcp_port: 8077
  # HTTP bearer token comes from this env var only — never from yaml.
  mcp_auth_token_env: MY_DAEMON_MCP_TOKEN

feedback:
  db_path: ./data/feedback.db

snapshot:
  # Bundles produced by `daemon snapshot create` land here. Each bundle freezes
  # the vector + graph + feedback state so the M3 analyzer and M4 observer LLM
  # can read a stable copy without contaminating live retrieval.
  dir: ./data/snapshots
  # `daemon snapshot prune` deletes bundles older than this; uncalled snapshots
  # never expire on their own.
  retention_days: 14

consolidation:
  # `daemon analyze <snapshot_id>` writes structural.json + weight_evolution.json
  # into <out_dir>/<snapshot_id>/. The M4 observer reads from the same place.
  out_dir: ./data/consolidation
  # Default lookback window for the hypothetical-weight-evolution replay.
  simulate_lookback_days: 7
  # Caps on the size of each ranked-list section of the report.
  max_communities: 8
  max_bridging_notes: 10
  max_bridge_edges: 20
  max_orphans: 20
  max_dangling: 20
  max_warm_edges: 20
  # Sampling for betweenness centrality. Full BC is O(VE), so we sample for
  # large graphs; 200 is a sensible default for personal vaults.
  betweenness_sample_k: 200

run:
  # `daemon run` — the foreground supervisor. Watches the vault, ingests after
  # things go quiet, and runs `daemon consolidate` overnight. Install it as a
  # background service with `daemon schedule install` (systemd --user on Linux,
  # a schtasks command on Windows).

  # Seconds of *quiet* before an incremental ingest fires. The timer restarts on
  # every new change, so a sync client rewriting a hundred notes produces one
  # ingest after it settles rather than one in the middle of it.
  debounce_seconds: 5

  # Local wall-clock time for the nightly consolidation, 24-hour "HH:MM".
  # Set to null to turn it off. It still obeys both writeback gates above
  # (agent.enabled AND agent.observer_enabled); with either closed the
  # supervisor logs why once and stops mentioning it.
  consolidate_at: "03:00"

  # Cadence of the "still alive" log line, in minutes. 0 silences it.
  heartbeat_minutes: 15

logging:
  level: INFO