Quickstart
A key, one request, then your client. About a minute.
Connect your client
Cursor, Claude Code, Codex, OpenCode and the OpenAI SDKs.
How memory works
Workspaces, the learning loop, and what you can read or erase.
API Reference
Every endpoint, with a playground.
What changes for your agent
- Learns from every turn. Corrections, repeated asks and wins turn into better practice. Your agent changes how it works, not only what it knows.
- Remembers across chats. Start a new chat and your agent already knows the decisions, preferences and gotchas from the last one.
- Picks up unfinished work. Open tasks carry over to the next chat, the next client and the next model, with why each one is still open.
- One memory per workspace. Each project learns on its own, so practice from one repo never leaks into another. Private to your account.
- Every major model. Hundreds of models from every major provider, behind one key, with activity, logs and limits.
The learning loop
1
Recall
Before it answers, your agent reads the notes that bear on the task: conventions, past decisions, what went wrong last time.
2
Act
Your model answers as usual, with the same tools and the same client.
3
Judge
Your next message shows how that went: a correction, a repeat, an error or a success. The agent judges it and how much it matters.
4
Improve
It rewrites the notes that led there, so the lesson sticks and the next turn, in any chat, starts better than the last.