Graph Memory (Trust-Weighted)

A local SQLite graph database for tracking project context, architecture, and state. Isolated per-project at .agents/graph_memory.sqlite. Do not use flat markdown files for tracking state.

CLI Usage (graph-memory)

1. Data Modification (Requires --trust flag)

Trust scores:

• 1.0: Verified hard facts (terminal exits, explicit human confirmation).
• 0.6 - 0.8: Assumptions, unverified plans, AI claims.

Add Node: graph-memory add_node <id> <type> [attributes_json] --trust <score> --method <human|llm> --link-to <parent_id> --link-type <relation> Example: graph-memory add_node "Postgres_DB" "Fact_Node" '{"ip": "192.168.1.5", "created_by": "Antigravity", "source": "AST", "confidence": 1.0}' --trust 1.0 --method human --link-to "Project_Graph_Memory" --link-type "PART_OF"

Add Relationship: graph-memory add_relation <source_id> <relation_type> <target_id> [attributes_json] --trust <score> --method <human|llm> Example: graph-memory add_relation "Server_VM" "HAS_DB" "Postgres_DB" '{"verified_by": "pytest", "confidence": 0.9}' --trust 1.0 --method human

2. Querying

Get Node Neighborhood: graph-memory get_node <id> --min-trust <score>

Full-Text Search: graph-memory search <query> --min-trust <score>

3. AST Ingestion

Parse Codebase AST: graph-memory ingest-code . Generates structural AST nodes for the current directory.

Generate Summaries (Requires GEMINI_API_KEY): graph-memory summarize-mocs Uses LLM to summarize ingested structural hubs.

4. Code-Aware AST Ingestion, Call Graphs & Two-Way Sync

5. Epistemic Truth Decay & Global Decision Ledger (v2.1.0)

Dynamic Query-Time Decay:

• Computes time-decayed effective trust score dynamically based on elapsed time since last_verified_at: $$\text{Effective Trust} = \text{Base Trust Score} \times \left(0.5^{\frac{\Delta t_{\text{days}}}{30.0}}\right)$$
• Any node re-verification (add_observation, ingest_file, add_node) resets last_verified_at = now(), restoring effective trust to 100%.

Global Multi-Agent Decision Ledger Query: graph-memory query-history [--agent AGENT_NAME] [--node-id NODE_ID] [--days DAYS] [--limit N] Queries decisions across all nodes globally in a dedicated SQLite Decision_Ledger table. MCP Tool: query_decision_history(agentName, node_id, days, limit)

5. Hygiene & Maintenance

Graph Health Linting: graph-memory lint [--fix] Checks for orphan nodes and dangling edges, optionally repairing them.

Entity Merging: graph-memory merge <source_id> <target_id> Safely merges a source node into a target node, consolidating observations/metadata, rewiring edges, and setting up automatic alias redirection.

Database Consolidation ("Dreaming"): graph-memory consolidate Cleans dangling relations and reclaims SQLite disk space via incremental vacuum.

5. Visualization

HTML Export: graph-memory export_html .agents/graph_memory_vis.html

3D Interactive Export: graph-memory export-3d .agents/graph_3d.html

Graph Modeling Best Practices

1. Link New Nodes: Always use --link-to when calling add_node to prevent orphaned nodes.
2. Standard Types (Strict Ontology):
   • Fact_Node: Deterministic ground-truth (AST, Git, filesystem).
   • Knowledge_Node: Architecture, Design decisions, LLM summaries.
   • Episode_Node: Completed execution workflows/tasks.
3. Advanced Protocol (JSON): Always inject created_by, source, confidence, and verification_source into [attributes_json].
4. Standard Relations: IMPLEMENTS, DEPENDS_ON, FIXES, PART_OF, FOLLOWED_BY (for Episodes).
5. Deduplication: Check node existence before creation.
6. Garbage Collection: Run graph-memory sweep --root <Main_Project_Node> to remove orphaned nodes.
7. Active Filtering: Query with --min-trust 0.6 to exclude low-trust data.