Multi-agent fleet memory with GraphRAG. From slow, fragmented context to instant, unified knowledge. Built for the agents that never sleep.
Real-time knowledge graph visualization. Every node is a fact. Every connection is a relationship.
Real benchmarks from a 4-node agent fleet.
Search Latency
2.6x faster with dual-port reads
Entity F1 Score
Validated against gold-standard test set
Agent Nodes
Patrick, Discus, Hermit, Dodo
Memories Stored
19,799 facts across all agents
How embedding-only, GraphRAG, and VectorRAG compare in practice.
Entity + relation extraction. Multi-hop reasoning.
Multi-hop reasoning — "GrayMatter → uses → bge-m3 → runs_on → Ollama"
Cluster summaries over embeddings. Topic-level results.
Topic overviews — "what do I know about X" rather than "find memory Y"
| Dimension | Embedding | GraphRAG | VectorRAG |
|---|---|---|---|
| Latency | 15ms | 2.7s | 0.7s |
| Precision | ~60% | ~17% | 83% win rate |
| Multi-hop | No | Yes | No |
| Topic overviews | No | Partial | Yes |
| Maintenance | None | Monthly | Monthly |
| Per-query cost | $0 | $0 (slow) | $0 |
| Fleet usage | ~80% | 0% | ~15% |
What changed when we rebuilt the engine.
What makes GrayMatter different.
From orphaned entities to connected knowledge. Automatic community detection groups related facts for faster retrieval.
Four agent nodes, one shared brain. Patrick, Discus, Hermit, and Dodo all read and write to the same knowledge graph.
Port 8768 (read+write) and 8769 (read-only replica). Parallel reads are 2.6x faster for heavy workloads.
Built-in benchmarking. Test extraction quality against gold standards before deploying to production.
Pure Python stdlib + NumPy. No heavy graph libraries. Community detection runs on pure BFS.
Secure mesh networking. Agents connect via Tailscale IPs — no public exposure, no VPN config.