Memory Expiry Rules for Superintelligent Agents: What Goes Stale
Assign an expiry date and a review trigger to each agent memory type, so prices, contacts and policies get rechecked before an agent relies on them.

Memory expiry for an AI agent means giving every stored fact a lifespan, a named owner and a review trigger, so the agent stops answering from prices, contacts or policies that have already changed. Set expiry by memory type rather than with one global setting: volatile facts such as prices get short windows and event-based triggers, while stable definitions get longer windows and scheduled reviews. An expired fact should never be silently trusted or silently deleted. It should be routed to a review, then corrected, replaced or removed.
Key takeaways
- Give each memory type its own expiry rule. A price list, a contact directory and a refund policy do not age at the same speed.
- Pair every time limit with a review trigger, such as a price update, a staff change or a published policy revision.
- When a fact expires, the agent should stop stating it as current and send the question to the record owner.
- Keep a visible queue of expired and overdue facts, reviewed by a named person on a fixed cadence.
- Test expiry as carefully as recall: ask questions whose correct answer has changed and confirm the agent stops repeating the old one.
Why agents keep outdated facts
A memory store records what was true when someone wrote it. Retrieval returns whatever matches the question best, and nothing in the store marks an item as old unless someone built that marking in. The agent then writes its answer with the same tone whether the source is from last week or from several years ago. That is why stale memory is more dangerous than missing memory. A missing fact prompts a question. A stale fact produces a confident answer that a customer or colleague may act on.
Most stale context comes from three sources: facts copied from documents that were later revised, contact details tied to people who changed roles, and prices or conditions that were valid only for a campaign or a season. The fix is not to write less memory. It is to attach a lifespan at the moment a fact is saved. For the broader problem, see how to stop stale context from misleading your AI agent.
Sort your memory by how fast it changes
Start with a short inventory that groups memory into four rough types. Each type gets its own expiry logic.
- Volatile facts: prices, stock levels, delivery windows, promotions and availability. These expire quickly and should be tied to the system or file that updates them.
- Relationship facts: contacts, account owners, supplier representatives and who approves what. These expire when the person or role changes, which usually means an event trigger from the directory or the team that maintains it.
- Working definitions: product descriptions, service names, internal terms and tone guidance. These expire slowly, with a scheduled review after each product or naming change.
- Policies and procedures: refund rules, warranty terms, approval thresholds and escalation paths. These expire on a revision date set by the policy owner, and any new version replaces the old one instead of sitting beside it.
Notice that the expiry for a policy is set by its owner, not by the memory system. A memory layer can hold the date, but it cannot decide the business rule. Mixed items deserve special care: a single note that combines a contact, a price and a policy should be split, because each part will expire on a different schedule.
A step-by-step plan to assign expiry and triggers
Work through the memory your first agent will actually use, not the whole company archive. A practical sequence looks like this:
- List every memory item the agent can read or write, with its source document, system or responsible person.
- Assign each item to one of the four types above and record a named owner who can confirm or retire it.
- Set an expiry window for each type, expressed as the maximum age the agent may rely on without a check.
- Attach at least one review trigger per item, such as a price file change, a role change in the directory or a new policy version.
- Write the agent's behavior for expired items: say that the information is being checked, give the owner's contact route, and do not quote the old value.
- Set a review cadence for the expired-items queue and decide who may extend, correct or delete each item.
- Run a short test set of questions whose answers changed in the last cycle before the agent goes live.
Which triggers work better than the calendar
Time windows are a backstop. Event triggers do most of the work because they fire when reality changes, not when a timer runs out. Useful triggers include:
- A source file or system record is updated, so the related memory items are marked for re-verification.
- A person listed as a contact leaves the role or changes responsibility.
- A new policy version is published in the document library.
- A staff member corrects the agent in conversation, which is a strong signal that neighboring items may also be stale.
- A supplier, carrier or partner issues a notice that changes conditions.
Illustrative scenario: a small distribution team stores a freight cutoff time in an agent's memory. The carrier changes its schedule, but the agent keeps quoting the old cutoff until the item's time window runs out. With a carrier-notice trigger attached to that item, the notice would have marked the fact for review on the day it arrived, and the agent would tell staff that the cutoff is being confirmed instead of repeating it.
Who signs off when a fact expires
Expiry needs a clear rule about authority. The agent can flag an item, lower its confidence and route the question to a person. It should not extend its own expiry, rewrite a policy or change a price in memory without an approval step. Permission review keeps that boundary visible: list who can write, who can extend and who can delete each memory type, then review that list on a schedule. Staff should also be able to see the expired-items queue, because the people closest to a customer often notice a change before any system does. For shared records, the rules for who may write are covered in shared memory write rules for teams.
How to test that expired facts stay retired
Testing is where expiry rules prove themselves. Build a small set of questions with known correct answers, and include several whose answers changed recently. For each question, check three things: the agent does not repeat the retired value, it says the item is being checked or points to the owner, and it names the current source when one exists. Failures belong in a log that the owner reviews, not in a chat thread that disappears.
Repeat the same test set on a schedule, and again after any change to the agent, the underlying model or the memory store. The OpenAI evals guide describes building repeatable evaluations of this kind, and the NIST AI Risk Management Framework treats monitoring and review as ongoing governance rather than a one-time setup.
How Be Human handles memory expiry in managed operations
Be Human designs expiry as part of the first engagement. The work starts by defining the one role the agent serves, the memory it needs and the owner for each memory type, with expiry windows and review triggers recorded beside every item. In managed operations, Be Human monitors the expired-items queue, runs recurring evaluations that include questions whose answers have changed, reviews the permissions that control who can write or extend memory, and makes controlled updates when a policy or source system changes. Be Human does not decide your prices or policies. Your named owners do, and the agent is configured to defer to them.
If you are not sure which memory in your company would do the most damage if it went stale, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
How long should an agent keep a price? Price memory should never be treated as a stable fact, so use a short maximum age plus a trigger on every price file or system update. If the source changes, the item is marked for review before the agent quotes it again. The right window depends on how often your prices move, and the operations owner should set it.
Should an agent ever delete an expired memory? Usually it should retire the item at expiry rather than delete it at once. A retired item can be removed from retrieval while the owner confirms the replacement, and kept in a change history for audit. Permanent deletion is a separate decision, governed by your retention rules and any policy on sensitive data.
Who decides the review cadence? The owner of each memory type decides, and the operations lead sets a common rhythm for the expired-items queue. A sensible starting point is a short regular check for volatile facts and a longer scheduled review for policies. The cadence should tighten when the same area produces repeated expiries.
What happens if an expired fact is the only answer available? The agent should say that the answer is being checked, give the route to the owner and avoid presenting the old value as current. If the question affects a customer decision, the handoff should happen in the same conversation rather than later. Offering the last known value with its recorded date and a clear warning is acceptable only when the owner has approved that behavior in advance.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


