What it does
MemPalace is a free, open-source memory system for AI agents. It keeps conversation history as verbatim text instead of summarising it, and finds relevant passages again with semantic search. The index is organised so searches can be scoped: people and projects become wings, topics become rooms, and the original content sits in drawers. It is local-first. The project says nothing leaves your machine unless you opt in, and that its core benchmark path needs no API key or LLM. ChromaDB is the default storage backend, with others available, including local SQLite-based ones, Milvus, Qdrant and pgvector. Embeddings can be computed locally or through an OpenAI-compatible server. You mine project files or saved coding-agent sessions into the store, then search them or load context at the start of a new session. The project ships a CLI, a 47-tool MCP server, a temporal knowledge graph kept in local SQLite, per-agent wings and diaries, and auto-save hooks for Claude Code, Codex CLI and Cursor. A Docker image can run the CLI or the MCP server, and skills help a coding agent install and operate it. The README publishes retrieval benchmark results with commands to reproduce them, and warns that its GitHub repository, PyPI package and documentation site are the only official sources. It is MIT licensed and requires Python 3.9 or later.
Capabilities
- Stores conversation history verbatim and retrieves it with semantic search
- Wings, rooms and drawers let searches be scoped to a person, project or topic
- CLI plus an MCP server with 47 tools
- Pluggable storage: ChromaDB by default, with SQLite, Milvus, Qdrant and pgvector options
- Temporal knowledge graph stored in local SQLite
- Auto-save hooks for Claude Code, Codex CLI and Cursor
When to use it
- Giving a coding agent long-term recall of earlier sessions and decisions
- Keeping agent memory on your own machine rather than in a cloud service
- Mining project files or saved transcripts into a searchable store
- Running a shared memory server for a team
