Nautilus Compass Memory — Claude Code MCP server: install
An MCP server for local long-term memory, session search, and drift checks for AI agents. Use it when Claude Code needs context across sessions.
claude mcp add nautilus-compass -- uvx nautilus-compassWhat Nautilus Compass Memory does
Nautilus Compass Memory provides local long-term memory for AI agents. It embeds raw text with BGE-m3, keeps data on the machine, and routes retrieval between utterance-level chunks and session-level search. The session mode combines BM25 with dense RRF. Ingest does not use an extraction LLM or a graph.
The server scores each prompt against an anchor set before the agent acts, which helps detect drift in the current instruction. It also includes cross-agent contracts, while the project description mentions byte-recomputable signed verification receipts. You can run it locally with uvx or connect over HTTP at https://compass.nautilus.social/mcp/.
The local setup uses uvx, so that command must be available for this option. The HTTP setup uses the remote MCP endpoint instead.
What it helps with
- retain agent memory across sessions
- retrieve facts from individual user turns
- check prompts for drift before an action
- coordinate several agents through contracts
Example prompts for the agent
- Save this project decision to long-term memory for later sessions.
- Search memory for the previous authentication fix in this project.
- Check my current prompt for drift against the stored rules.
How to install Nautilus Compass Memory
1Platform CLI
claude mcp add nautilus-compass -- uvx nautilus-compassuvx runs the command named after the package; if the README names another one, use uvx --from nautilus-compass <command>
2Platform CLI: remote server
claude mcp add --transport http nautilus-compass-memory https://compass.nautilus.social/mcp/1Platform CLI: remote server
codex mcp add nautilus-compass-memory --url https://compass.nautilus.social/mcp/2Platform CLI
codex mcp add nautilus-compass-memory -- uvx nautilus-compassuvx runs the command named after the package; if the README names another one, use uvx --from nautilus-compass <command>
Config file
{
"mcpServers": {
"nautilus-compass-memory": {
"url": "https://compass.nautilus.social/mcp/"
}
}
}{
"mcpServers": {
"nautilus-compass-memory": {
"command": "uvx",
"args": [
"nautilus-compass"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from nautilus-compass <command>
Config file
{
"servers": {
"nautilus-compass-memory": {
"type": "http",
"url": "https://compass.nautilus.social/mcp/"
}
}
}{
"servers": {
"nautilus-compass-memory": {
"type": "stdio",
"command": "uvx",
"args": [
"nautilus-compass"
]
}
}
}uvx runs the command named after the package; if the README names another one, use uvx --from nautilus-compass <command>
1Platform CLI: remote server
gemini mcp add -t http nautilus-compass-memory https://compass.nautilus.social/mcp/2Platform CLI
gemini mcp add nautilus-compass-memory uvx nautilus-compassuvx runs the command named after the package; if the README names another one, use uvx --from nautilus-compass <command>
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 Nautilus Compass Memory
How do I install Nautilus Compass Memory in Claude Code?
For a local setup, run claude mcp add nautilus-compass -- uvx nautilus-compass. This registers the MCP server through uvx.
What does Nautilus Compass Memory do?
It gives AI agents local long-term memory with session retrieval. It also checks prompts for drift and supports contracts between agents.
What is the difference between the local and HTTP modes?
The local mode runs through uvx and keeps memory on the machine. The HTTP mode connects to https://compass.nautilus.social/mcp/.
