What an agent should remember, and what it should forget

A memory that keeps everything is a transcript. The value is in the choosing.

A test for what to keep

Before a fact is written, it should pass four questions.

1. Was it stated? Keep what the user said. Do not keep what the model guessed.
2. Is it stable? “Is vegetarian” lasts. “Is hungry” does not.
3. Will it be useful later? Would knowing this change a future answer?
4. Is it allowed? Not a secret, not a sensitive category you have no basis to hold.

Most sentences in most conversations fail at least one. An extractor that returns nothing for an ordinary chat is working properly.

Write notes that stand alone

A note is read months later with no surrounding conversation. “She prefers the second one” is useless. “Maya prefers aisle seats” is not. Include the subject, use plain statements, and attach the date.

Handling contradiction

People change jobs, cities and minds. When a new note arrives:

• find the most similar existing notes;
• if one says the same thing, do nothing;
• if one conflicts, replace it and keep the newer date;
• if it is unclear which is right, keep both with dates and let the agent ask.

Never leave two flat contradictions in the store. The model will not reliably pick the newer one.

Four ways to forget

On request. The user says “forget that”. Delete the note, and anything derived from it. This one is not optional.

By replacement. A new fact supersedes an old one, as above.

By expiry. Some notes have a natural end: a trip, a deadline, a temporary address. Store an expiry date when you write them.

By decay. Notes that have not been recalled for a long time are ranked lower and eventually archived. This keeps recall sharp without anyone making a decision per note.

A size check

Look at the memory of your heaviest user. If it reads like a tidy profile, the rules work. If it reads like a diary, tighten the extractor.