config.yaml

# narv - Global Configuration

api:
  # Daily limit for LiteLLM API. Triggers budget exhaustion safeguards when reached.
  litellm_budget_daily: 1000
  retry_base_delay_sec: 1
  max_retries: 3

  # LiteLLM endpoint configuration
  # Set a custom API base URL (e.g., "http://localhost:4000" for LiteLLM Proxy).
  # Leave empty to use litellm's built-in provider routing.
  litellm_api_base: ""
  # Separate endpoint for embedding models (optional). Falls back to litellm_api_base if empty.
  litellm_embedding_api_base: ""
  # Model names follow litellm naming convention (e.g., "gemini/gemini-2.0-flash", "anthropic/claude-sonnet-4-20250514").
  # See: https://docs.litellm.ai/docs/providers
  model_fast: "gemini/gemini-3.1-flash-lite"
  model_slow: "gemini/gemini-3.1-flash-lite"
  model_embed: "gemini/gemini-embedding-001"

cognitive:
  # Routing threshold. Events with urgency >= this value bypass the slow model 
  # and are routed to the fast model. Lower this to prioritize speed and reduce cost.
  urgency_threshold_fast: 0.8

  # Load threshold. If cognitive load exceeds this, the system may force a rest phase 
  # or throttle inputs. Increase to sustain high-intensity processing longer.
  cognitive_load_max: 0.8

  # Token limits for different cognitive phases. 
  # Tune these to balance response detail against API latency and token budget.
  fast_tokens: 1024
  slow_tokens: 4096
  dmn_tokens: 4096
  sleep_tokens: 4096
  # Number of session events passed to Reflection prompt.
  # Reduced from hardcoded 30 to limit prompt token growth.
  reflection_window: 15

system:
  log_level: "INFO"

  # Controls how quickly the system enters the Sleep phase (memory consolidation).
  # Decrease to consolidate memories more frequently; increase to stay reactive longer.
  idle_cycles_for_sleep: 60
  max_cognition_failures: 3
  max_resume_loop: 10
  cycle_interval_seconds: 50.0

  # Capacity of Working Memory (Redis). When exceeded, older session events are pruned 
  # to a core set (e.g., 20). Increase to retain a longer raw event history before pruning.
  session_memory_keep_count: 200

  # LLM Attention Focus Size. Defines how many recent events are injected into the 
  # cognitive prompt. Tune to balance context richness against token cost and noise.
  context_window_size: 40

perceptor:
  # Urgency weights assigned to different types of perceptual inputs.
  # Adjust these to control how easily specific events trigger 'fast' cognitive routing.
  urgency_anomaly: 0.9
  urgency_normal: 0.2
  urgency_shell_output: 0.3

actor:
  sandbox_root: "data/sandbox"
  allowed_commands: "mkdir,rm,mv,grep,echo,ls,cat,pwd,sed,awk,python,python3"
  timeout_ms: 30000

memory:
  redis_url: "redis://localhost:6379/0"
  chroma_persist_dir: "./data/chroma"
  neo4j_uri: "bolt://localhost:7687"
  neo4j_user: "neo4j"
  neo4j_password: "narv_memory_2026"

  # Similarity threshold for vector search. Increase for stricter, more precise memory recall;
  # decrease to allow broader, fuzzier associations.
  cosine_threshold: 0.75
  top_n_default: 5

  # --- Capacity Management ---
  # Redis (SessionMemory) limits.
  redis_max_entries: 500
  redis_ttl_seconds: 86400       # 24 hours

  # ChromaDB (LongTermMemory) capacity and eviction rules.
  chroma_max_entries: 10000
  chroma_eviction_batch: 500

  # Importance threshold for Long-Term Memory. During capacity management, 
  # entries with importance below this value are permanently forgotten.
  chroma_eviction_threshold: 0.7 # Marks importance < 0.7 for eviction

  # Neo4j (StructuralMemory) capacity and compression rules.
  neo4j_max_nodes: 5000

  # Triggers graph compression when node count exceeds this threshold.
  neo4j_archive_threshold: 4000

  # Importance threshold for Structural Memory. Nodes below this value 
  # are archived/compressed to maintain graph traversal performance.
  neo4j_min_importance: 0.3      # Marks nodes with importance < 0.3 for compression