Mnemos MCP
MCP server with a memory of the agent's own: it decides what to keep and wakes with it each session, stored in one SQLite file
- 158 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- mnemos connect --write changes the agent configuration and installs hooks
- Background jobs run every four hours and at night, and on Claude they spend model time
- The memory accumulates facts about you and the project in a file on disk
Install
- Works in
Manual install
pip install "mnemos-continuity[mcp,embeddings]"After installing, run mnemos embeddings download, then mnemos connect claude-code --write.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- mnemos connect --write changes the agent configuration and installs hooks
- Background jobs run every four hours and at night, and on Claude they spend model time
- The memory accumulates facts about you and the project in a file on disk
This is third-party code. Review the repository files before installing.
What it does
Mnemos gives an AI agent long-term memory that the agent keeps itself, in its own words. It stores a conclusion about the user, a joint decision or its own mistake, and at the start of each session gets those notes back along with a note it left for itself. There are twelve tools: keep, recall, correct what turned out wrong and answer a question that the memory itself asks. Memory follows how brains work. An entry carries three measures: strength, stability and accessibility. Entries fade along a forgetting curve without being deleted: a weak entry goes to an archive and can return on recall. Each recall strengthens an entry and links it to the ones that came with it. When a detail has nearly gone, the agent is asked what it taught and the lesson is kept. Everything sits in one SQLite file on your machine. Upkeep runs every four hours and a deeper pass at night. The mnemos log command opens a local page with the agent's days, its notes and everything it holds. If you uninstall Mnemos the file stays: deleting it is left to you.
Who it is for. People who want an agent to remember a project, decisions and corrections across sessions with no external service.
Good fit when
- The agent gets the same correction a second time and should remember it
- Decisions made together need to survive between sessions
- Several agents, for example Claude Code and Codex, should share one memory
- You want to see and correct what the agent remembered on a local page
Not a fit when
- You need an index of your documents or code search: Mnemos is not built for that
- You need a shared team knowledge base: the memory belongs to the agent and sits on one machine
- You want to write the memory yourself: the agent writes the entries and your words stay yours
Example request
Connect Mnemos to Claude Code, keep the conclusion from today's decision about the migration and bring it back in the next sessionMCP
- Transport
- stdio
- Authentication
- not required
Limitations
Python 3.10 or newer is required. The embeddings extra pulls in PyTorch (several hundred MB), and the embedding model downloads separately (about 90 MB); without it recall matches words only. The scheduler and night jobs run on macOS and Linux, while on Windows the memory and tools work. Surfacing memories per message currently works only in Claude Code. The project is early, and the author confirms it. Keep the installed package where it was set up: hooks call mnemos by its full path.
How to disable. mnemos uninstall shows what it would remove and mnemos uninstall --write does it. The memory file stays on disk, so delete it by hand if you no longer need it.
FAQ
Which agents work with Mnemos?
Any MCP client. The README has mnemos connect commands for Claude Code, Codex, Grok, Gemini CLI, Hermes and OpenClaw, and mnemos mcp install for Claude Desktop and Cursor.
Where is the memory stored?
In one SQLite file on your machine. No company sits in the chain, so the memory cannot be switched off remotely, and uninstalling Mnemos does not erase the file.
How does Mnemos differ from a plain notes store?
Entries fade along a forgetting curve, strengthen on recall and link to each other, and the agent writes them itself and gets them at session start. By the author's own test plain files do about as well on facts, but they do not get the agent to keep notes or bring them back at start.
What is needed to install it?
Python 3.10 or newer and pip install "mnemos-continuity[mcp,embeddings]". The embeddings extra installs PyTorch (several hundred MB), and mnemos embeddings download fetches a model of about 90 MB once. Without it recall matches words.
Does Mnemos run on Windows?
The memory and tools work, but scheduled jobs (upkeep every four hours and the night pass) need macOS or Linux.
Discussion
No questions or reviews yet.
Related
MCP serversLocal memory with evidence: what an agent learned and where from, and when a source is corrected the dependent records are retired
A long-term memory MCP server for AI agents: facts fade along a forgetting curve and related records are compressed into a summary
Long-term memory layer for AI agents: decisions and facts are stored as JSON next to the code and found again in a later session.
A self-hosted MCP server for long-term agent memory: structured records, versioned rollback and a visual dashboard
An MCP server for long-term agent memory: compresses dialogues into atomic facts, resolves pronouns and anchors relative time to absolute dates
One local binary that gives Claude Code, Codex and Cursor shared memory across sessions, with a knowledge graph on disk and no API keys
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Mnemos (this card) | no | not required | 158 | |
| MCP Memory Service | no | not required | 2,117 | a shared service memory for several agents |
| OMEGA Memory | no | not required | 225 | memory focused on searching accumulated data |
| agentmemory | yes | not required | 30,314 | a memory store without forgetting or night upkeep |
README in short
The README opens with the idea of continuity for the agent: its memory in its own words, in one file on the owner's machine. It continues with a pip quick start and a table of connect commands, a feature list, the neuroscience-based memory model and a day schedule with upkeep every four hours and a night pass.