Teaching Memento

Memento gets better at your library the more you correct it — and every lesson it learns is a sentence you approved, sitting in your instructions where you can read, edit or delete it.

Thumbs up, thumbs down

Answers in the assistant, changes in Activity and the analysis Memento shows on a capture in your Inbox each carry a thumbs up and a thumbs down. A thumbs-down opens a small box for a note — “receipts belong under Finance”, “that’s the wrong Anna” — which is optional, and usually the most useful part. Tap the other thumb to change your mind; tap the same one again to edit the note or remove your rating.

Your notes are private to you. Memento’s operator view counts thumbs by model and by kind of output, and never shows what anyone wrote.

Corrections become suggested rules

Memento also notices the corrections you make by doing rather than voting: a change you reject or undo, or a capture where you picked a different category or person than it suggested. When a few of those have piled up and you have stopped for a while, it looks for a pattern — at most once a day for each space — and may suggest a short standing rule, such as “File receipts under Finance, not Topics.”

Suggestions appear at the top of Settings → Instructions, with the corrections they came from. For each one you can:

• Add to instructions — exactly as written;
• Edit — reword it first, and your wording is what gets added;
• Dismiss — nothing happens and it goes away.

Suggest now asks for suggestions straight away instead of waiting; it counts towards the same once-a-day limit. Looking for a pattern uses your connected AI account, like an inline edit does, but it is not an agent run and does not use up your daily runs.

What an added rule does

An added rule is appended to that space’s instructions under Rules learned from your feedback, and from the next run on everything the agent does there follows it — filing a capture, running a routine, answering you in chat. It is ordinary text in your instructions: edit or delete it there whenever you like. See Spaces & tuning for how instructions work.

What Memento will not learn

Your notes never reach the agent as you wrote them, and nothing you or a captured document says becomes a rule without you adding it. Memento also refuses to add a rule — and tells you why — when it would:

• break the page format or the checklist grammar behind your task list;
• contradict its own safety rules: citing sources, keeping inferred facts marked as inferred, never silently deleting, asking before risky changes, staying inside this space, never sending anything anywhere;
• take your instructions past their 10,000-character limit;
• repeat something your instructions already say.

If that is a false alarm, reword the rule, or write it into your instructions yourself — there the agent follows it and points out any conflict in its answer instead.