Memocat MCP Server

MemoCat MCP Server — Self-Hosted AI Agent Memory and Vector RAG

MemoCat is a self-hosted MCP memory server for Claude Desktop, Cursor, ChatGPT integrations, and autonomous AI agents. It gives MCP clients private, persistent, semantically searchable long-term memory. memocat-mcp is a Model Context Protocol server for Montycat, an all-in-one, self-hosted data and memory engine for AI agents.

Montycat combines the database, vector search, on-device embeddings, persistent and in-memory storage, real-time subscriptions, and data governance in one engine. MemoCat exposes those capabilities through MCP, so an agent does not need a separate vector database, embedding API, event broker, or governance service to build persistent memory and retrieval-augmented generation (RAG).

Memories are embedded on-device and recalled by meaning, metadata, timestamp, or exact key. No cloud vector database, external embedding API, or per-query bill.

Features

• Self-hosted long-term memory for MCP-compatible AI agents.
• Semantic vector search with metadata and time-range filtering for RAG.
• Persistent memory, in-memory working spaces, bulk writes, updates, and deletion.
• Real-time memory-change subscriptions without database polling.
• Multi-agent scopes, shared memory, delegated-owner governance, and policy explanations.
• Keyspace lifecycle, semantic-model controls, snapshots, and revocation-safe watch buffers.
• One-command uvx memocat-mcp entry point with native/Docker engine bootstrap.

Why

LLM agents forget context between runs. MemoCat stores each fact in Montycat, embeds and indexes it automatically, and retrieves relevant memories through 22 agent-readable MCP tools. It is both a retrieval layer for RAG and a durable memory layer for autonomous agents. Unlike a collection of separate database, vector, embedding, messaging, and policy services, Montycat provides the full memory stack as one all-in-one engine that can run under your control.

Install MemoCat MCP

The fastest option is uvx, which runs the latest published package in an isolated environment:

uvx memocat-mcp

If uvx is not installed yet:

curl -LsSf https://astral.sh/uv/install.sh | sh
uvx memocat-mcp

For a persistent command-line installation, use pipx:

pipx install memocat-mcp
memocat-mcp

You can also install it into an existing Python environment:

python -m pip install memocat-mcp
memocat-mcp

MemoCat requires Python 3.10 or newer. The package is published as memocat-mcp on PyPI.

Quick start with a Montycat engine

MemoCat reuses a configured Montycat Semantic engine or attempts the supported native/Docker bootstrap path. For an existing engine:

export MONTYCAT_URI="montycat://memory-agent:password@localhost:21210/memories"
uvx memocat-mcp

Tools

Real-time memory watch

Other memory servers can only be polled: ask again, and again, in case something changed. Montycat has native live subscriptions, so this one pushes.

agent B: memocat_await_memory_change(scope="shared", timeout_sec=60)
                    ⏳ sleeps — no polling, no wasted tokens
agent A: memocat_remember({"text": "the deploy key rotated"}, scope="shared")
agent B: ← returns in milliseconds with the key, the value, and the event

Two agents, one shared scope, one notices what the other just learned. Pass the returned next_seq back as since_seq to resume exactly where you left off — changes that happen between calls are buffered, not lost.

Memory namespaces are also exposed as MCP resources (memocat://memory/<keyspace>) with resources.subscribe support, so clients that implement resource subscriptions get notifications/resources/updated pushed to them as well. Both surfaces share one engine subscription.

Subscriptions open on demand and close when idle (MONTYCAT_WATCH_IDLE_TIMEOUT), so users who never watch pay nothing.

Montycat Semantic engine requirements

• Python 3.10+ when installing MemoCat through uv, pipx, or pip.

• Access to a Montycat Semantic engine. uvx memocat-mcp first reuses an existing engine, then attempts the supported native/platform installation path, and finally falls back to Docker. Semantic search is enabled by default in the Semantic edition.

  To start the engine manually with Docker, pick the tag for your CPU—the tag carries the architecture:

  Apple Silicon (M1/M2/M3/M4) — use arm64-semantic:

  docker run -d --name montycat -p 21210:21210 -p 21211:21211 \
    -e MONTYCAT_SUPEROWNER="admin" -e MONTYCAT_PASSWORD="change-me" \
    -v montycat_data:/var/lib/.montycat \
    montygovernance/montycat:arm64-semantic
  

  Intel / AMD (x86_64) — use semantic:

  docker run -d --name montycat -p 21210:21210 -p 21211:21211 \
    -e MONTYCAT_SUPEROWNER="admin" -e MONTYCAT_PASSWORD="change-me" \
    -v montycat_data:/var/lib/.montycat \
    montygovernance/montycat:semantic
  

  On Apple Silicon the plain semantic tag is the amd64 image and runs under emulation, where the embedding runtime's warm-up crashes. Use arm64-semantic — a native build, not a workaround. Unsure which you have? uname -m prints arm64 on Apple Silicon and x86_64 on Intel.

  Port 21211 is the subscription server and is required for memocat_await_memory_change (real-time watch); without it the other tools still work.

Docker Compose deployment

Use Compose when you want a reproducible local deployment with a persistent Semantic engine and an MCP container on the same private Docker network. Docker is optional when you already manage a reachable Montycat server.

Create a .env file beside compose.yaml:

MONTYCAT_USERNAME=admin
MONTYCAT_PASSWORD=replace-with-a-strong-password
MONTYCAT_STORE=memories
MEMOCAT_VERSION=0.4.2
# Apple Silicon: arm64-semantic. Intel/AMD64: semantic.
MONTYCAT_IMAGE_TAG=semantic

Start the engine and build the MCP image:

docker compose up -d montycat
docker compose build mcp

After the public Docker image is released, use it instead of building from source:

docker pull montygovernance/memocat-mcp:0.4.2

The Compose service uses montygovernance/memocat-mcp:${MEMOCAT_VERSION} and waits for the Semantic engine health check before launching MCP. The image runs as an unprivileged memocat user and supports both AMD64 and ARM64.

The image installs the released montycat>=1.2.2,<2 Python client declared in the package metadata.

The engine data is stored in the named montycat_data volume. Ports 21210 and 21211 are published for debugging and external clients; the MCP container uses the private montycat:21210 network address. Credentials are passed as separate environment variables, so passwords with URL-special characters need no URL encoding. Port 21211 carries live subscription traffic for memocat_await_memory_change.

MCP uses stdio, so do not run it as a web service. Configure a desktop MCP client to invoke the Compose service on demand:

{
  "mcpServers": {
    "memocat": {
      "command": "docker",
      "args": [
        "compose",
        "-f", "/absolute/path/to/memocat-mcp/compose.yaml",
        "run", "--rm", "-T", "mcp"
      ]
    }
  }
}

For Apple Silicon set MONTYCAT_IMAGE_TAG=arm64-semantic in .env; the plain semantic image is AMD64. Stop the stack with docker compose down; include -v only when you intentionally want to erase persisted memories.

Engine auto-start

MemoCat first reuses an engine already reachable through MONTYCAT_URI or the host/port settings. If none is running, it looks for an installed montycat_bin and then attempts the official platform route:

MemoCat asks the shared Montycat release catalog for the current Semantic artifact for macOS or Windows. Artifact URLs are treated as opaque, and the package's adjacent .sha256 is required and verified before Installer opens; verified packages are cached by filename. If catalog discovery is unavailable, automatic native installation falls through to Docker instead of silently installing an older package. Override the URL with MEMOCAT_INSTALLER_URL, pin a release with MEMOCAT_ENGINE_VERSION, or adjust the Installer completion budget with MEMOCAT_INSTALLER_TIMEOUT. On Linux, set MEMOCAT_APT_INSTALL_COMMAND to use an organization-managed mirror or package command. ARM64 Linux goes directly to Docker because the official APT repository is AMD64-only. Set MEMOCAT_AUTOSTART=off to disable all installation/start attempts.

Connect Claude Desktop

Add MemoCat to claude_desktop_config.json, then restart Claude Desktop:

{
  "mcpServers": {
    "memocat": {
      "command": "uvx",
      "args": ["memocat-mcp"],
      "env": {
        "MONTYCAT_URI": "montycat://memory-agent:agent-password@localhost:21210/mystore"
      }
    }
  }
}

Connect Cursor

Add the same server definition to your Cursor MCP configuration:

{
  "mcpServers": {
    "memocat": {
      "command": "uvx",
      "args": ["memocat-mcp"],
      "env": {
        "MONTYCAT_URI": "montycat://memory-agent:agent-password@localhost:21210/mystore"
      }
    }
  }
}

Connect OpenAI Codex

Codex can register the local stdio server directly from a terminal:

codex mcp add memocat \
  --env MONTYCAT_URI="montycat://memory-agent:agent-password@localhost:21210/mystore" \
  -- uvx memocat-mcp

Confirm the registration with codex mcp list, then start a new Codex session.

ChatGPT integration

MemoCat currently runs as a local stdio MCP server. It works directly with clients that can launch local MCP commands, including Claude Desktop, Cursor, and Codex. A ChatGPT connector requires a remotely reachable MCP transport and cannot connect directly to this stdio command. Remote HTTP transport is not included in the current package; do not expose the engine's database port as an MCP endpoint.

Connect to a remote TLS engine

Keep the normal montycat:// connection URI and enable TLS separately:

export MONTYCAT_URI="montycat://memory-agent:agent-password@db.example.com:21210/mystore"
export MONTYCAT_TLS=true
uvx memocat-mcp

For a desktop client, add "MONTYCAT_TLS": "true" beside MONTYCAT_URI in the server's env object. The remote engine must present a certificate trusted by the machine running MemoCat. Setting MONTYCAT_URI disables local engine auto-install and auto-start because it explicitly selects a managed engine.

Security and delegated-owner setup

Use a delegated Montycat owner such as memory-agent for the MCP process. Grant that owner only the keyspace read/write and provisioning capabilities its agent needs. Keep the superowner credential in a separate bootstrap or governance-administration workflow.

The read-only memocat_policy_view, memocat_policy_explain, and memocat_policy_history tools expose policy information for the authenticated owner. They do not accept an owner override and cannot grant, revoke, deny, or otherwise mutate policy. When automatic keyspace provisioning fails, MemoCat also requests a read-only policy explanation and appends it to the original engine error when available.

memocat_remove_keyspace is destructive and remains engine-authorized. A delegated owner may remove a keyspace through creator authority or an explicit remove-keyspace grant unless policy contains an overriding denial. MemoCat closes active watches and releases resource subscriptions before requesting removal.

Semantic management is always keyspace-scoped. The MCP server does not expose database-wide semantic controls; Montycat checks manage-semantic, creator authority, denials, and model allow-lists for every enable or disable request.

Snapshot tools are likewise keyspace-scoped and work only with in-memory keyspaces. MemoCat does not expose the global snapshot-rate setting. A Snapshot rate is not set response means scheduling has not been configured on the engine; it is distinct from a governance denial.

Active watches use short authorization leases because the current engine checks read authority when a subscription opens but does not terminate that connection after a later revocation. MemoCat revalidates against the engine's filtered structure view, closes the subscription on access loss, removes MCP resource ownership, wakes pending callers with an error, and permanently purges buffered changes so they cannot be replayed after access is restored.

Configuration

memocat_create_keyspace works with delegated-owner credentials when policy grants provision-keyspace for the requested store, storage type, and semantic model. The store must already exist for delegated owners. With superowner credentials, creating the first keyspace also creates a missing configured store in the same engine request. memocat_forget deletes one record and requires write authority for its keyspace. The engine makes every final authorization decision.

Memory scoping (multi-tenant)

Pass scope (an owner/user id) to any memory tool to isolate that owner's memory. Because Montycat's semantic search runs per keyspace, each scope gets its own keyspace mem_<scope> — so semantic recall for one owner never sees another owner's memories:

remember(value={"fact": "..."}, scope="alice")      # -> keyspace mem_alice
semantic_search(query="...", scope="alice")          # searches only mem_alice
remember(value={"fact": "..."}, scope="shared")      # -> the shared keyspace

• Per-owner private memory — scope="<owner>" → mem_<owner>, auto-created on first use when the configured owner has provisioning authority.
• Shared/common memory — scope="shared" → the MONTYCAT_SHARED_KEYSPACE.
• Group memory — use a group id as the scope (e.g. scope="team_eng").
• Single-tenant — set MONTYCAT_SCOPE once and omit scope per call.

This maps onto Montycat's keyspace governance. In production, run one server instance per agent or service with delegated-owner credentials and grant only the provisioning and data authority it needs.

Isolation note: scope is routing convenience, not authenticated identity. With one server instance sharing one connection, scopes provide logical keyspace organization. For credential-enforced isolation, run one server instance per owner with that owner's delegated credentials; the engine then denies cross-owner access. Reserve superowner credentials for bootstrap and governance administration.

Links

• MemoCat MCP on PyPI
• MemoCat MCP source
• Report an issue
• Changelog
• Montycat documentation
• Montycat Semantic engine on Docker Hub
• Montycat Python client on PyPI

License

MIT.