Local-first semantic memory for AI agents — conversations become atomic memories in a SQLite knowledge graph, exposed through FastAPI and MCP.
PythonSQLiteLanceDBMCPMemory
Overview
Atlas turns conversations into atomic memories and stores entities, relationships, and temporal context as a knowledge graph in SQLite. Retrieval is hybrid over LanceDB vectors and SQLite full-text search, with results reranked by recency, access-based reinforcement, and memory decay. The Python SDK is the core; a FastAPI service, CLI, MCP server with 9 tools, and a React + Three.js web app are optional adapters over it.
Highlights
- Extracts atomic memories, entities, and relationships from conversations
- SQLite knowledge graph as the source of truth, with temporal context on every memory
- Hybrid retrieval in LanceDB, reranked by recency, access-based reinforcement, and decay
- Local Transformers embeddings, so indexing and search stay on-device
- 9 MCP tools plus a FastAPI service and CLI over the same Python SDK
- 3D Three.js view of the memory graph
Stack
- Python 3.11+
- Memory SDK, workers, and CLI
- SQLite
- Source-of-truth knowledge graph
- LanceDB
- Local vector index
- Transformers
- On-device embeddings
- FastAPI
- HTTP service
- MCP
- Memory tools for agent clients
- React + Three.js
- 3D memory graph web app