config.yaml

# HAIM Configuration — Phase 4.5
# All hardcoded constants are centralized here.

haim:
  version: "4.5"
  dimensionality: 16384  # 2^14, must be multiple of 64

  # Vector encoding
  encoding:
    mode: "binary"  # "binary" (Phase 3.0+) or "float" (legacy)
    token_method: "bundle"  # "bundle" (XOR+permute) or "hash"

  # Memory tier thresholds
  tiers:
    hot:
      max_memories: 3000
      ltp_threshold_min: 0.7
      eviction_policy: "lru"

    warm:
      max_memories: 100000
      ltp_threshold_min: 0.3
      consolidation_interval_hours: 1
      storage_backend: "mmap"  # "mmap" (Phase 3.0) or "qdrant" (Phase 3.5)

    cold:
      storage_backend: "filesystem"  # "filesystem" or "s3"
      compression: "gzip"
      archive_threshold_days: 30

  # LTP (Long-Term Potentiation) decay parameters
  ltp:
    initial_importance: 0.5
    decay_lambda: 0.01  # Exponential decay rate
    permanence_threshold: 0.95  # Above this, memory is immune to decay
    half_life_days: 30.0  # For synaptic connections

  # Hysteresis (prevent boundary thrashing between tiers)
  hysteresis:
    promote_delta: 0.15  # LTP must exceed threshold by this much to promote
    demote_delta: 0.10   # LTP must fall below threshold by this much to demote

  # Redis (Phase 3.5)
  redis:
    url: "redis://localhost:6379/0"
    stream_key: "haim:subconscious"
    max_connections: 10
    socket_timeout: 5

  # Qdrant (Phase 3.5)
  qdrant:
    url: "http://localhost:6333"
    collection_hot: "haim_hot"
    collection_warm: "haim_warm"
    binary_quantization: true
    always_ram: true
    hnsw_m: 16
    hnsw_ef_construct: 100

  # GPU (Phase 3.5)
  gpu:
    enabled: false
    device: "cuda:0"
    batch_size: 1000
    fallback_to_cpu: true

  # Observability (Phase 3.5)
  observability:
    metrics_port: 9090
    log_level: "INFO"
    structured_logging: true

  # Persistence paths
  paths:
    data_dir: "./data"
    memory_file: "./data/memory.jsonl"
    codebook_file: "./data/codebook.json"
    concepts_file: "./data/concepts.json"
    synapses_file: "./data/synapses.json"
    warm_mmap_dir: "./data/warm_tier"
    cold_archive_dir: "./data/cold_archive"

  # Security (Phase 3.5.1)
  security:
    # api_key: "mnemocore-beta-key"  # <--- REMOVED: Must be set via HAIM_API_KEY env var or here explicitly

  # MCP (Model Context Protocol) bridge
  mcp:
    enabled: true
    transport: "stdio"  # "stdio" recommended for local MCP clients
    host: "127.0.0.1"
    port: 8110
    api_base_url: "http://localhost:8100"
    timeout_seconds: 15
    allow_tools:
      - "memory_store"
      - "memory_query"
      - "memory_get"
      - "memory_delete"
      - "memory_stats"
      - "memory_health"

  # Dream Loop (Subconscious background processing)
  dream_loop:
    enabled: true
    frequency_seconds: 60  # Seconds between dream cycles
    batch_size: 10  # Number of memories to process per cycle
    max_iterations: 0  # Maximum iterations (0 = unlimited)
    subconscious_queue_maxlen: 10000  # Max queued IDs (null/0 = unlimited)
    ollama_url: "http://localhost:11434/api/generate"
    model: "gemma3:1b"

  # Phase 4.0+: Semantic Consolidation
  consolidation:
    enabled: true
    interval_seconds: 3600  # 1 hour between consolidation cycles
    similarity_threshold: 0.85  # Hamming similarity threshold (0.85 = 15% distance)
    min_cluster_size: 2  # Minimum cluster size for merging
    hot_tier_enabled: true  # Consolidate HOT tier
    warm_tier_enabled: true  # Consolidate WARM tier

  # Phase 4.1: XOR-based Project Isolation
  attention_masking:
    enabled: true  # Enable/disable project-based memory isolation

  # =========================================================================
  # Subconscious AI - BETA FEATURE
  # =========================================================================
  # This is a BETA feature that enables autonomous background AI processing
  # for memory management, dream synthesis, and micro-self-improvement.
  #
  # WARNING: This feature is experimental and may change without notice.
  # Must be explicitly enabled by setting 'enabled: true'.
  # All safety defaults are conservative - review before enabling in production.
  # =========================================================================
  subconscious_ai:
    # BETA FEATURE - Must be explicitly enabled
    enabled: true
    beta_mode: true

    # Model configuration
    model_provider: "ollama"  # ollama | lm_studio | openai_api | anthropic_api
    model_name: "phi3.5:latest"
    model_url: "http://localhost:11434"
    # api_key: null  # For API providers
    # api_base_url: null

    # Pulse configuration
    pulse_interval_seconds: 120
    pulse_backoff_enabled: true
    pulse_backoff_max_seconds: 600

    # Resource management
    max_cpu_percent: 30.0
    cycle_timeout_seconds: 120
    rate_limit_per_hour: 50

    # Operations
    memory_sorting_enabled: true
    enhanced_dreaming_enabled: true
    micro_self_improvement_enabled: true  # Initially disabled

    # Safety
    dry_run: false
    log_all_decisions: true
    audit_trail_path: "./data/subconscious_audit.jsonl"
    max_memories_per_cycle: 10