Afair

You re-explain yourself to every AI you open. Your working style, your project names, your family — every tool starts fresh, and you pay the tax in repeated context-setting. afair addresses this by acting as a passive memory store that connects to your AI tools over MCP. As you have conversations, afair extracts relevant facts, ranks them, resolves entities across sessions, and keeps a structured vault. The next tool you open reads from that vault automatically. You never manually file anything.

The vault’s architecture is the differentiating design choice. It is single-tenant by definition — one machine per user, never a shared database. Encryption runs at the database layer via SQLCipher and at the file layer via AES-256-GCM, and decryption requires your key. The vendor states the structure afair builds is human-readable data you can open, inspect, and correct yourself. Nothing is hidden in model weights. When the extraction logic gets something wrong, you fix it. The full vault exports to a single file at any time.

This fits one profile well: an individual switching between Claude, ChatGPT, Cursor, Codex, or Perplexity who wants personal context to persist without vendor lock-in. It breaks down quickly outside that profile. There is no API, so any application that needs to read your memory programmatically rather than through an MCP client is blocked. There is no multi-user or team memory model — the architecture is explicitly single-tenant. The managed hosted option the vendor describes is not yet live.

On the integration side, the vendor states afair connects to Claude.ai, ChatGPT, Perplexity, Claude Code, Cursor, Codex, and Warp through MCP. Three models from Anthropic, OpenAI, and Google are used internally for extraction and ranking — meaning the self-hosted version still calls external model APIs rather than running inference locally. Teams expecting a fully air-gapped deployment need to account for that dependency.