Stop repeating yourself to your AI.

The context tax

AI tools are brilliant for the first ten minutes of a conversation and amnesiac after it ends. What your team decided, how your setup works, which client wants what — none of it survives the chat.

Per-user memory features don't fix the team problem. Your assistant slowly learns you — and still knows nothing about the decision your colleague made yesterday. Ten teammates, ten quietly diverging memories.

One memory, every tool

Rafter is one MCP server holding the team's memory: decisions, conventions, gotchas, client facts. Every connected tool — Claude, ChatGPT, Cursor, Codex — reads the same memory, and writes new learnings back as the team works.

Memories link to each other and to the skills that use them, so a fetch returns not just a note but what it cites and what cites it. Ask what the team decided about pricing and the answer arrives with its reasons attached.

Everything is versioned, and re-creating something a human deleted requires an explicit confirm. Agents write — the team stays in charge.

What you’d actually type

• “What do we know about [a client or project]?”Answered from the team's actual notes, with what they cite.
• “Save this as a memory: [the decision you just made].”Ten seconds now; the next person doesn't re-litigate it.
• “Why do we [rule]? Check Rafter before guessing.”The reasons live next to the rule — the graph keeps them attached.