[ docs · concepts ]

Verified-outcome memory

Grist remembers — but only what was proven. A lesson enters memory when its outcome is verified: the tests passed, you approved the work, or you corrected it. Everything else is forgotten on purpose.

Why verified only

An agent that remembers everything remembers its mistakes with the same confidence as its wins, and poisons future sessions with confident nonsense. Grist's rule is stricter: memory is a record of verified outcomes, not a transcript. What a session merely claimed is never stored. What it proved is.

A run Verified? Memory Next run,briefed unverified work is not stored
admission is the whole trick: proof in, everything else out

The outcome kinds

OutcomeMeaning
tests_passedThe change was verified by the project's own checks.
user_approvedYou reviewed the work and approved it.
user_correctedYou fixed the agent's work — the correction itself is the lesson, and it's weighted heavily.
check_failedA cautionary record: this approach failed its check, so the next run doesn't repeat it.

How it behaves

Where it lives

By default, memory is a file in the project (.grist/memory.json) — plain JSON you can read, diff, and commit. Teams can run the shared store so a whole team's verified outcomes brief every member's sessions. The desktop app ships a memory browser: search what Grist remembers about a project, and forget any record with one click.

Memory never updates model weights. It is retrieval plus feedback — which is exactly why you can inspect it, correct it, and delete it.