Superintelligence Memory Hygiene: Monthly Checks for Stale Facts
A monthly routine to find outdated prices, contacts and policies in AI agent memory, assign a named owner to each fact, and retire it with a change log.

AI agent memory hygiene is a recurring review that finds facts the agent still treats as true but no longer are, such as old prices, retired contacts and superseded policies. Run it monthly on a short list of high-risk records, check each fact against a named current source, and retire every outdated fact under a named owner. The aim is not a perfect memory but one where every stored fact can be traced to a source and to a person who answers for it.
Key takeaways
- A stale fact does more damage than a missing one, because the agent states it with the same confidence as a correct fact.
- Review a short list of high-risk record types every month: prices, contacts, policies and the limits the agent may act within.
- Every remembered fact needs a source, a date of last verification and a named owner.
- Retire a fact by correcting it, expiring it or removing it, and log which choice you made and why.
- Managed operations can run the sweep and bring exceptions to the owner, but the owner decides what is true.
Why stale facts cost more than missing facts
An agent that does not know a price will usually say so or ask a question. An agent that remembers an old price will state it fluently, in the same tone it uses for correct answers, and staff may not notice until a customer does. Staleness is hard to see because the memory looks healthy: the record exists, it matches the question and it was once true.
The most common carriers of stale facts are fee schedules, contact details for people and vendors, and policies that changed in a meeting or an email rather than in a document. Each of these changes in a different place, and rarely triggers a memory update on its own. For the wider problem of context that outlives its usefulness, see how to stop stale context from misleading your AI agent.
Which record types to check every month
Not every fact needs a monthly look. Start with the record types where a wrong answer creates a commitment, a refund, a misdirected message or a compliance problem. A short list is more likely to be reviewed honestly than a long one.
- Prices, discounts, fee schedules and payment terms
- Contact points for people, teams, suppliers and escalation paths
- Policies on returns, cancellations, service levels, warranties and approvals
- Limits the agent may act on, such as who can approve an exception
- Product or service descriptions that changed with a recent release
- Status, deadlines and decision makers for open projects
Everything outside these groups can move to a quarterly pass or to an event-driven check. Use the agent's own logs to see which facts it cited most often in the last month, because those deserve review first.
The monthly sweep, step by step
Keep the same order each month so that results can be compared from one sweep to the next. The sweep is mechanical, which makes it a good task for an agent to prepare and a person to finish.
- Step one: export the high-risk records with their source, last verified date and owner fields.
- Step two: sort by last verified date, oldest first, and flag every record older than the review interval set for its type.
- Step three: compare each flagged fact with the current source of truth, such as the price list, the policy document or the contact directory.
- Step four: mark each fact as confirmed, changed or unknown. Unknown is a valid result and goes to the owner rather than being guessed.
- Step five: for each changed fact, write the replacement value and the source that supports it, then send it to the owner for approval.
- Step six: ask the agent five to ten questions that staff really ask, and check that its answers match the confirmed facts.
- Step seven: record the sweep in a change log with the date, the reviewer, the facts confirmed and the facts changed.
The sweep proves nothing if its results are not written down. The change log is the evidence.
Assign a named owner to every fact
A fact without an owner is checked by nobody. An owner is one person or one role that answers for whether the fact is true, not whoever happened to type it into memory. Finance may own fee schedules, the operations lead may own escalation contacts and the support manager may own return policy. Write the owner into the record itself so that the agent and the team see the same name.
Owners do not need to run the sweep. They need to confirm changes and accept or reject proposed replacements within a set window. If an owner leaves or changes role, the records they held move to a reassignment queue rather than remaining under a name that no longer applies. For rules on who may write to shared records, see shared memory write rules for teams.
How to retire a fact without losing its history
Retiring a fact means one of three things. Correct it when the value changed and the record still describes the same thing. Expire it when it should stop being used after a date, such as a time-limited offer. Remove it when it should never have been stored, for example because it came from an unverified message.
- Correct: replace the value, keep the record identifier, and log the old value and the reason for the change.
- Expire: set an end date, stop the agent from using the record after that date, and keep it for audit.
- Remove: delete the record from active memory and confirm that copies in summaries or caches are cleared too.
Do not overwrite silently. A staff member who sees an answer quoted from last quarter should be able to trace what the agent knew and when it knew it. For expiry rules by record type, see memory expiry rules for superintelligent agents, and for the deletion side, see how to delete AI agent memory.
Illustrative scenario: a support agent quotes a retired fee
Illustrative scenario: A support agent for a small equipment rental business holds a record of its late return fee. The fee was changed after a staff meeting and the price sheet was updated, but the agent's memory still held the earlier amount. During the monthly sweep, the operations lead compares the record with the current price sheet, marks it as changed, and the support manager approves the replacement. The change log shows the old value, the new value, the source document and the date of approval.
The next customer question returns the current fee. If someone later asks why an earlier answer differed, the log explains it. Without the sweep, the first sign of the problem would likely have been a customer disputing a charge.
What a managed routine looks like with Be Human
Be Human treats memory hygiene as part of managed operations rather than a one-time cleanup. Monitoring compares recorded facts with the sources named in each record and reports the ones past their review interval. Recurring evaluations run the questions staff actually ask and flag answers that disagree with confirmed facts. Permission review checks that each owner still exists and that no agent can change a fact it does not own. Changes to prices, contacts or policies go to the named owner before they are written, and every write is logged. The agent proposes, and the people who own the facts decide.
If you are unsure which of your records would cause the most trouble if it went stale, the free Be Human AI diagnostic is a practical way to find the first task worth handing to an agent.
Frequently asked questions
How often should a business run a memory hygiene check? Monthly is a sensible default for high-risk records such as prices, contacts and policies, because those change often enough to cause errors within a quarter. Lower-risk records, such as background descriptions, can move to a quarterly pass. Trigger an extra check whenever a price list, a policy document or a team structure changes, since those events are when facts most often go stale between sweeps.
Who should own the monthly check itself? The sweep needs one accountable person, usually an operations lead or a knowledge owner, who schedules it, signs off the change log and escalates unresolved items. Individual facts still belong to their own owners, such as finance for fees or support for policies. The sweep owner coordinates, and the fact owners decide what is true.
Can an agent run the check on its own? An agent can do much of the mechanical work: list records past their review date, compare them with a named source and draft proposed replacements. It should not approve its own changes or alter prices, contacts or policies without the owner's confirmation. Write that boundary down and check it during each permission review.
What if a fact cannot be verified during the sweep? Mark it as unknown rather than guessing, and stop the agent from quoting it as current. Give the owner a fixed window to respond, and if none arrives, reduce the fact's visibility until someone confirms it or removes it. An unverified fact left in place is still a stale fact.
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.


