Embeds persistent memory into Claude with three backing stores: SQLite for local dev, pgvector for semantic search, and Neo4j for graph traversal across entities and relationships. You call `add` when new information arrives and `recall` to pull ~200 tokens of context per query, regardless of conversation length. The retrieval pipeline is deterministic and keeps token usage flat while full-context replay grows quadratically. Includes bi-temporal tracking so you can replay any past state and an auditable supersession chain for multi-agent environments. Self-hosted, runs as a library or MCP server, designed for teams that need tenant-isolated memory with compliance-grade audit trails. One-command quickstart wires Postgres, Pinecone Local, and Neo4j in Docker.
claude mcp add --transport stdio io.github.bolnet-memwright uvx memwright