memory-arbiter-mcp — Claude Code MCP server: install
A local SQLite MCP server for agent memory, traceable facts, version history, and conflict notices. Use it when memory must stay on the machine.
Open on GitHubWhat memory-arbiter-mcp does
memory-arbiter-mcp gives coding agents a local fact store over MCP. It keeps the original text, provenance fields, timestamps, and an append-only edit history. Full-text search and vector indexes support recall, while its four product tools expose memory operations to the client.
Embeddings run through a local GGUF model. An optional mDeBERTa checkpoint can act as a local CPU conflict judge; it reports competing facts without editing them or selecting a winner. The vector setup uses memory-arbiter-mcp[vec,semantic-local], followed by mema setup --install and mema doctor.
The server needs explicit client and agent_id values in its configuration or launch environment. It refuses to start when either identity is blank, so this setting belongs in the setup checklist.
What it helps with
- store agent facts in a local SQLite database
- retrieve prior context with text and vector search
- keep provenance and revisions for each memory
- record conflicts between competing facts
- protect user-confirmed memories from silent edits
Example prompts for the agent
- Remember this project fact and attach its source details.
- Find memories related to this task and show the evidence behind the answer.
- Check these new facts against stored memories for conflicts.
How to install memory-arbiter-mcp
Install
The MCP registry has no launch command for it: see the repository README
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about memory-arbiter-mcp
How do I install memory-arbiter-mcp in Claude Code?
Run pip install "memory-arbiter-mcp[vec,semantic-local]", then mema setup --install, and finish with mema doctor. Set client and agent_id in the configuration afterward.
What does memory-arbiter-mcp do?
It stores agent facts in local SQLite, retrieves them through text and vector indexes, preserves revision history, and reports conflicts. State-changing governance actions require authorization for that specific action.
What is required for it to start?
The configuration or launch environment must provide both client and agent_id. A blank value for either one prevents startup; embedding models run locally.
