Crew Memory Graph
A multi-agent collaborative memory system built on the FalkorDB graph database. It uses a shared graph structure to enable memory across sessions and precise handoffs between agents.
Timeline (last 90 days)
Lusavio built the Crew Memory Graph project at the Graph Hacks hackathon.
Crew Memory Graph contains three agents: Planner, Researcher, and Executor.
Crew Memory Graph's memory is divided into three layers: episodic memory, semantic memory, and procedural memory.
Crew Memory Graph uses a separate FalkorDB graph instance per tenant to achieve tenant isolation.
There is no LLM in Crew Memory Graph's query path; a deterministic pattern-based planner compiles questions into graph traversals.
In the released repository, 25 tests pass, and 2 live FalkorDB integration tests that require a configured server are skipped.
The frozen benchmark of 14 questions (8 of which are multi-hop, scored by evidence path length) scored 14/14.
The application runs live on a real FalkorDB Cloud instance, and at the most recent check the graph contained 206 nodes and 616 edges.
Sending 10 questions simultaneously to the live demo, all 10 were answered, with a median answer time of about 21 seconds on free-tier hosting.
In graph-based handoff, the handoff itself is a node that points via edges such as REFERENCES to specific entities and events in memory, and the new agent obtains precise raw information by traversing the graph.
1 more timeline events
All Facts (11)
Lusavio built the Crew Memory Graph project at the Graph Hacks hackathon.
50%UnverifiedCrew Memory Graph contains three agents: Planner, Researcher, and Executor.
50%UnverifiedCrew Memory Graph's memory is divided into three layers: episodic memory, semantic memory, and procedural memory.
50%UnverifiedCrew Memory Graph uses a separate FalkorDB graph instance per tenant to achieve tenant isolation.
50%UnverifiedThere is no LLM in Crew Memory Graph's query path; a deterministic pattern-based planner compiles questions into graph traversals.
50%UnverifiedIn the released repository, 25 tests pass, and 2 live FalkorDB integration tests that require a configured server are skipped.
50%UnverifiedThe frozen benchmark of 14 questions (8 of which are multi-hop, scored by evidence path length) scored 14/14.
50%UnverifiedThe application runs live on a real FalkorDB Cloud instance, and at the most recent check the graph contained 206 nodes and 616 edges.
50%UnverifiedSending 10 questions simultaneously to the live demo, all 10 were answered, with a median answer time of about 21 seconds on free-tier hosting.
50%UnverifiedIn graph-based handoff, the handoff itself is a node that points via edges such as REFERENCES to specific entities and events in memory, and the new agent obtains precise raw information by traversing the graph.
50%UnverifiedThe repository also includes an in-memory development backend labeled as holding no deployment state; the live deployment runs only FalkorDB, with no agent holding state in process memory.
50%