magnemo

Magnemo is a Governed Memory tool for AI agents, developed by Silver Valley Technologies Inc., with the slogan "Sessions end. Memory doesn't." It keeps AI memory on the user's own computer, stored as plain text files. The AI writes down what it learns during work, but nothing becomes official memory until the user approves it; every memory shows who wrote it and who approved it. When the agent writes something, Magnemo asks the user in chat, and the user can approve it with one click.

Key capabilities and governance mechanisms

• Approval gate: What the agent writes only enters a staging area, and is never written directly to official storage. The AI gets four MCP tools—retrieve, stage, bootpack, handoff—but "promotion/approval" is not among them; approval authority always belongs to the user, and the user's answer does not pass through the agent.
• Provenance file format: Every note is a Markdown file with provenance frontmatter, with fields including author, written, source, status, reviewed_by, supersedes, strength. Provenance itself is the file format.
• Recall guarantee (phase one—governed recall): Search returns only promoted notes; staged notes do not participate in search; the boot pack shows staged material and marks it as staged, including pending thread titles, top staged note titles, and the truncated line of the last handoff plus a pointer to its note.
• Partition isolation: Partition isolation is enforced per agent (MAGNEMO_PARTITIONS), and cross-partition access is treated as DENIED.
• Supersession: Supersession is explicit; old notes are archived with forward links, and nothing is deleted.
• Rejection records: Rejections are retained and recorded—rejections can also provide lessons.
• Trust ledger: Every promotion/rejection is written to an append-only Trust Ledger, including the operator and reason. Trust is computed from that record, not subjectively asserted.
• Local and privacy: Memory is kept in plain files on the user's computer and can be opened by the user; the package imports no network modules, and every run of magnemo doctorScan check; the only network traffic is a git push performed when adding a backup destination.

Usage Installation is a one-paste three-line command, or let the agent install it itself: open Claude Code in any project folder and paste the corresponding instructions (its magnemo mount step writes to .mcp.json, the file Claude Code reads). In Cursor, Windsurf, or other clients, use the configuration block for that client. After restarting, tell the agent "Use Magnemo to remember what this project is", the agent writes a line, Magnemo asks for approval in the chat, and after one-click approval you can ask it what it remembers. The local core is an MCP server (uvx magnemo-mcp, with MAGNEMO_VAULT pointing to the folder where memories are stored, which can be created with magnemo init ./vault) and a command-line tool (review, promote, reject). One server behaves consistently across clients. It has been tested on Claude Code (a session can mount the server, list the four tools and call them, and raise the approval question in the chat); the configuration blocks for clients such as Claude Desktop, Cursor, VS Code/Copilot, Windsurf, Cline, Continue, Gemini CLI, Codex CLI, and Zed are all provided according to their respective official documentation, but are officially marked as untested. Remote entry points (ChatGPT and hosted clients) are planned for version 0.7.0.

Applicable scenarios and limitations Suitable for developers and teams who want AI agents to accumulate long-term memory locally in an auditable, approvable way. Memories can coexist with any existing memory system (note folders, Obsidian, RAG stack). It uses Python 3.11+, has zero dependencies, and is Apache-2.0 licensed. The project is currently in Open Beta, and the official guidance is to read KNOWN_LIMITS.md before entrusting it with anything that must not be lost. If magnemo cannot be found after installation, it is usually because the pip scripts directory is not in PATH; using python3 -m magnemo always works. Information about fees, pricing, and similar details is not currently provided.