Montycat MCP — Shared AI Memory README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Montycat MCP — Shared AI Memory listing page.
Montycat MCP - Shared Memory for AI Agents
A self-hosted MCP server that gives AI agents persistent, searchable memory. Claude, Codex, Cursor, and any Model Context Protocol client write to one memory and read each other's.
- Memory that survives the chat. Decisions, preferences, and project context carry into the next session.
- One memory, many agents. Every MCP client you use works from the same facts.
- Recall by meaning, keyword, or both. Vector search finds a memory when the wording differs, BM25 nails exact identifiers, and hybrid mode fuses the two. Exact-key and metadata lookup too.
- Yours. Server, engine, and embeddings run on your machine. No hosted memory service, no cloud embedding API.
Install
Claude Desktop — download
montycat-mcp.mcpb
and drag it into Claude Desktop. No Python needed. That link always serves the
current release; every release
also carries a version-named copy and a .sha256 to check it against.
Claude Code — install uv, then:
/mcp confirms the montycat server is connected.
Codex, Cursor, other MCP clients — point your client's stdio config at
uvx montycat-mcp (Python 3.10+). For Codex:
The engine
Memory lives in a Montycat Semantic engine. Montycat MCP starts a local one for you, so most people can stop reading here.
Point it at an engine you already run:
Or start one yourself with Docker:
On Apple Silicon use the arm64-semantic tag instead — semantic is the amd64
image, and it crashes under emulation. Port 21211 carries live memory watches.
Use it
Talk to your agent normally; it picks the tool.
Remember that the team chose PostgreSQL for the billing service.
What did we decide about the billing database?
Save this to the shared
engineeringscope.
scope decides where a memory lives — alice for private, engineering for a
team, shared for common. It is a namespace, not a security boundary: for real
isolation, give each MCP server its own least-privilege Montycat credential.
Tools
| Need | Tools |
|---|---|
| Store | montycat_remember, montycat_remember_bulk, montycat_update, montycat_forget |
| Recall | montycat_semantic_search, montycat_recall, montycat_list_memories |
| Inspect schemas | montycat_list_enforced_schemas — check field names and data types before structured writes or filtered retrieval |
| Collaborate | montycat_await_memory_change — wait for another agent's write, no polling |
| Namespaces | montycat_list_keyspaces, montycat_create_keyspace, montycat_remove_keyspace |
| Admin | semantic index, snapshot, and policy tools — see the plugin guide |
Destructive tools are declared as such, so your client's confirmation prompts apply.
Configuration
| Variable | Purpose |
|---|---|
MONTYCAT_URI | Connection string: montycat://user:password@host:port/store |
MONTYCAT_TLS | true for a remote TLS engine |
MONTYCAT_DEFAULT_KEYSPACE | Memory namespace; memory by default |
MONTYCAT_SCOPE | Default scope when a call omits one |
MONTYCAT_AUTO_PROVISION | Create a permitted scope on first use; true by default |
MONTYCAT_AUTOSTART | off to require an already-running engine |
Compose setup and the full variable list: compose.yaml and the plugin guide.
More
Changelog · Privacy · Issues · Docs · Docker Hub
Existing MemoCat installs keep working: memocat-mcp, MEMOCAT_*, and
memocat:// are still supported. New setups should use the Montycat names.
MIT