Here is a conversation that happens about eighteen months into every successful AI deployment, and almost never before it.
Someone discovers that a decision made in July was based on a wrong figure. The figure has since been corrected. The question on the table is not what is the number — that is settled — but what did we believe when we decided, and when did we find out otherwise.
In most systems this question has no answer, because the correction destroyed the evidence. The record says what is true now. What it said in July is gone, and with it the ability to tell the difference between a bad decision and a reasonable decision made on bad information.
Those two things have very different consequences for a person's career, an insurance claim, or a regulator's letter.
Overwriting is a choice that looks like a default
When a fact changes, the obvious thing is to update the row. It is what databases are for. It keeps the store small and the reads simple, and it is what nearly every memory layer does, usually without ever describing it as a decision.
But it quietly asserts something false: that the current state is the whole truth. A customer's address is not "42 Mill Lane". It is "14 Dock Street until March 2026, then 42 Mill Lane". The second form answers questions about invoices sent in February. The first form cannot, and will confidently mislead you if you ask.
Two clocks, not one
The fix is well understood in database theory and strangely rare in AI products: record two timelines per fact.
When it held in the world. The contract ran from January to December. The address was current from March.
When you believed it. You learned the renewal moved on 18 June. You corrected it on 2 October.
With both, three questions become answerable that are otherwise reconstructions: what did we believe on 31 July, how did this deal change over the quarter, and what changed while I was away. And one question becomes answerable that matters more than all of them: was the decision reasonable given what was known at the time.
Keeping the superseded fact rather than deleting it is what makes this work. A correction should record that something stopped being current on a given day, not erase the fact that it was ever believed.
Why this is now a procurement question
For a few years AI memory was a product concern. It is becoming a compliance one, for three reasons that are arriving together.
Auditability. If an AI system contributed to a decision with financial or personal consequences, the reasonable question is what it was working from. "We no longer hold that" is a poor answer in a dispute and a worse one in a regulated industry.
Erasure. The right to be forgotten requires knowing everything held about a person across every store, deleting it, and proving you did. A system that scatters derived facts without provenance cannot honour that request, no matter how sincerely it intends to — it cannot even enumerate what it holds.
Provenance. "The AI said so" is not a source. A fact that carries the system, the record id and the link it came from can be checked, challenged and corrected by the person who owns that system. A fact that carries nothing can only be believed or ignored.
The companies that will handle the first serious audit of an AI-assisted decision are the ones that treated memory as a record rather than a cache — and you cannot retrofit a history you did not keep.
What to ask a vendor
Four questions, and they are short enough to put in an email:
- When a fact changes, what happens to the old one?
- Can I read the memory as it stood on a given date?
- For any fact, can you show me the system, record and link it came from?
- If someone asks to be forgotten, can you find everything held about them, delete it everywhere, and give me a receipt?
The answers to these are architectural. They are decided early, they are expensive to change, and a vendor who has not thought about them will usually tell you so within a sentence of being asked.