What your agent should forget

Forgetting is a feature

Persistent memory is what turns a chatbot into an operator: your agent remembers your policies, your tone, and last week’s decision without being told twice. But memory that only ever grows quietly fills with expired offers, old prices, and private details that stopped being useful weeks ago. Treat forgetting as a deliberate part of the design, not neglect.

The test for any saved item is simple: would a future decision be worse if this were outdated and the agent used it anyway? A current refund policy earns its place. A one-off instruction from a rushed Tuesday does not. How to judge it: if you cannot name a future task the memory improves, it is clutter — and clutter is where stale advice hides.

The trade-off is real. Forget too eagerly and you drop useful context, so the agent asks you the same question twice. That is why forgetting needs a policy, not a mood: keep what guides real decisions, and put an expiry or a review date on everything that will change.

Delete these by default

Some memories carry risk the moment they are stored and add nothing to a future decision. Delete these by default, without waiting for a review:

• One-time secrets — passwords, verification codes, and temporary access links, worthless after use and dangerous if retained.
• Raw private conversations — once the outcome you needed is recorded, the full transcript is liability, not memory.
• Unapproved drafts — opinions or replies the agent wrote but you never signed off on.
• Superseded instructions — the old version of any rule you have since changed.
• Sensitive data outside an active workflow — personal or financial details no current job needs.

How to judge it: if an item is both risky to hold and unlikely to improve a decision next month, it goes. The failure to avoid is deleting something a live workflow still relies on — a policy the agent quotes daily, for instance. When you are unsure whether a memory is truly finished, expire it rather than delete it: set a review date and let the next cleanup confirm.

Expire time-bound information

Most memory is neither permanent nor disposable — it is simply true for a while. Campaign dates, prices, team assignments, project status, and policies each change on their own schedule. Store the fact together with a review date and a source the agent can re-check, so a claim can be verified instead of trusted forever.

How to judge it: a healthy time-bound memory can answer “says who, and until when?” If it cannot, the agent should ask you rather than guess. The failure case is a silent expiry: a price rose in March, the memory still says the old number, and the agent quotes it with total confidence because nothing ever told it the fact had aged.

The trade-off is upkeep — review dates only help if someone returns to them. A recurring monthly review, run as a scheduled task, keeps the habit from depending on your memory instead of the agent’s. Automatic per-item expiry dates are on the eeky AI roadmap and marked “Coming soon”; until then, the review date lives in the note and your monthly pass enforces it.

Separate delete from correct

Two different problems get mixed up here. If a fact is wrong, correct it — replace it with the accurate version and record where the new fact came from. If a fact is merely finished — a closed project, an ended promotion — remove it. Deleting a wrong fact without replacing it leaves a gap the agent may fill by guessing; keeping a corrected fact beside its old version leaves two truths and no way to know which is current.

How to judge it: after any edit, there should be exactly one current answer to the question the memory addresses. The classic failure is the competing-versions trap — an old shipping fee and a new one both sit in memory, and retrieval surfaces whichever it happens to match first. Overwrite; do not append.

The trade-off is that overwriting loses history. When the previous value might matter later — a price you may need to prove you once charged — record the change in activity history rather than keeping a second live memory the agent could act on by mistake.

Run a monthly memory cleanup

A policy no one enforces decays. Once a month, read the agent’s memory the way you would review a filing cabinet shared with a fast, literal colleague. Start with the memories that drive external actions — the ones behind messages, prices, and offers customers see — because that is where a stale fact does visible damage. Then work through the oldest unsourced facts and anything marked sensitive.

Give every item one of four verdicts: keep, correct, expire, or delete. How to judge the pass: a good cleanup ends with fewer items than it started, and every survivor can say what it is for and when it should be checked again.

The trade-off is time — a thorough review costs perhaps twenty minutes. The failure it prevents is the expensive one: an agent confidently acting on something that was true in January. If twenty minutes feels like too much, that is usually a sign the memory has grown without a forgetting habit for too long.