Afair

Every AI you use remembers a different version of you. What one learns is gone when you open the next, so you repeat the same decisions, relationships, preferences, and history at the start of another session.

afair gives them one shared vault. It connects through the Model Context Protocol (MCP), so the same memory is available in Claude, your coding agent, and whatever you use next.

A growing memory usually becomes another thing to maintain. afair takes that work away. Background agents read what your tools record, discover what belongs together, keep the useful understanding current, and let old noise fade from attention. The structure follows your life instead of a fixed template. You do not file, tag, or curate it.

The result stays visible. You can read what afair believes, follow it back to the source, and correct it when it gets you wrong. Corrections become part of the history instead of silently replacing it.

What goes in is your whole life: your work, the people you love, and the personal things you'd rather an AI just knew.

The data is append-only and yours. afair is open source and single-tenant, and you can export all of it whenever you want. The name is short for "as far as I remember", the hedge it's built to make unnecessary.

Two ways to use it

Run it yourself. This repository is the whole thing, AGPLv3. Self-host it on your own machine or server and you own every layer end to end. It is free, forever. The quickstart is below.

Or let afair.ai run it for you. Managed hosting, with your own isolated EU instance, backups, export, and updates handled for you, is coming soon. Join the early-access list at afair.ai.

Same code either way. The hosted product is one deployment of this repo, not a separate proprietary fork.

The three commands

afair exposes exactly three tools, and they are frozen for good:

• remember stores something durable: a decision, a person who matters, a date you can't miss, a preference.
• recall pulls back what is relevant to the moment.
• observe logs what the AI just did, so the vault keeps up.

Once you hand your AI the short setup snippet, it calls these on its own. Nothing reaches the vault unless a call puts it there.

Run it yourself

Requires Python 3.12+ and uv.

uv sync
cp .env.example .env
# add ANTHROPIC_API_KEY, or a key for any other provider (the model is yours to pick)
uv run python -m afair

The server comes up on http://127.0.0.1:8765. Point a CLI or desktop client (Claude Code, Codex, Cursor, GitHub Copilot) at it, connect, and you are done. Web clients that run in the cloud (Claude.ai, ChatGPT) instead need a public HTTPS deployment and a quick OAuth setup, covered in docs/self-hosting.md. Every environment variable is documented inline in .env.example; per-client connection guides live in docs/clients.

Install it with your coding agent

Already in Claude Code, Codex, or another coding agent? Hand it the prompt below and it sets itself up: clone, dependencies, config, and the MCP wiring, then it proves the round-trip.

Set up afair (the open-source MCP memory server at https://github.com/afairai/afair) on this machine and connect this tool to it. Clone it, run uv sync, copy .env.example to .env, and ask me for an LLM provider API key to put there. Then run uv run python scripts/install_clients.py to wire my MCP clients and add the afair instruction snippet. Tell me the command to start the server (uv run python -m afair) and how to keep it running. Once it is up, remember a test fact and recall it to prove it works. Follow docs/self-hosting.md and docs/clients/; put my API key only in .env, nowhere else.

That runs locally and serves CLI and desktop clients out of the box. To reach web clients (Claude.ai, ChatGPT), follow the public-deployment notes in docs/self-hosting.md.

Teach your AI to use it

Connecting the server is half of it; the other half is making your AI reach for the tools on its own. Paste the short instruction snippet into the client's persistent instructions (CLAUDE.md, AGENTS.md, Custom Instructions, or .cursorrules), and it will recall context at the start of a conversation, remember what's durable, and observe what it does, without you prompting it each time. The same snippet works for every client. For Claude Code, Codex, and Cursor, scripts/install_clients.py writes both the connection config and the snippet for you; for GitHub Copilot it writes the connection config and prints the one per-repo snippet step (Copilot reads instructions per workspace).

Works with

Claude Code, Claude.ai, ChatGPT, Codex CLI, Cursor, Windsurf, Copilot, and anything else that speaks MCP over Streamable HTTP.

Architecture

Four layers, one source of truth:

• Substrate. Append-only SQLite with FTS5 and sqlite-vec, content-addressed. The log is never rewritten.
• Interpretation. Versioned views built over the substrate. Regenerate them without touching a single stored event.
• MCP surface. Versioned and additive. A signature that has shipped keeps working.
• Agents. Background workers extract evidence, discover what belongs together, and maintain cited living syntheses. Cluster names and boundaries can change as the evidence changes. The user never defines a category.

The complete design, and the eight invariants that hold it together, live in VISION.md. Start there if you want the why.

Documentation

Contributing

Pull requests are welcome. Read CONTRIBUTING.md for setup, the checks that must pass, and the invariants a change cannot break. Found a security issue? See SECURITY.md, and please report it privately.

License

afair is released under the GNU Affero General Public License v3.0 (LICENSE). You can self-host it, fork it, and modify it freely. If you run a modified version as a network service for others, you publish your changes under the same license. The hosted offering at afair.ai is one deployment of this code, not a separate proprietary fork.

In one line: free to use, free to host yourself, share back if you run it as a service for others.

Made in Germany

Built in Germany. The hosted instances run in the EU, under EU jurisdiction.