Context at your cursor.
You are about to send. The brain recalls in the right partition, and only a strong hit earns one sentence. Nothing is inserted until you press Tab.
Situation in. One line out.
Addressee, subject, last sentence, enough to recall under the right branch. The small model writes only from those facts. Empty when the hit is weak, with a reason you can read.
- Recall first, no free-form invention
- Strong hit only; otherwise silence
- Tab to accept, never auto-insert
- Outcome and send events stay in the ledger
The same rows your tools get.
MCP recall carries which branch you are reading. Remember names its undo. Account endpoints list what was written, rows, not estimates.
Session brief and tools see the brain before the first call.
Paste the MCP URL; sign in; the directory profile is ready.
SDK remember / recall / ask on the same project key.
Promises caught on the way out.
When a send carries a promise, the memory it writes is a plan, inferred-from-send, held until it is kept. That is how the cursor path stays honest about what you said you would do.
- Detector is deterministic, not a second brain
- A plan stays held until it is kept or dropped
- Works with branches so the debt sits on the right deal