Answers
Isn't out-of-date AI memory worse than having no memory at all?
The risk is real and it is the right thing to ask about: an assistant that states a dead rule with confidence does more damage than one that asks. What decides it is whether anything ends a memory other than someone remembering to delete it. Here every rule and decision is written together with the condition that ends it — the revocation, the date, the event that makes it false. When that condition is met the memory is retired and stops being served, and a decision that replaces an earlier one names the one it replaced instead of sitting beside it. The failure you are describing belongs to stores that only accumulate. Expiry is part of the record, not a cleanup someone has to schedule.
Last updated September 15, 2026
A memory that cannot die is the one that goes bad
Write-only is the normal shape of this problem. Context goes in, retrieval comes out, and nothing ever leaves, so the store grows and the oldest entries become indistinguishable from the newest in everything but a timestamp nobody reads. Every entry here is written with what would kill it: an explicit revocation, a date it stops being true, or an event that forces a review. An entry with no such condition is not accepted — which means there is no way to add something that quietly lives forever.
A replacement names what it replaces
The second half of going stale is contradiction: two entries that were each right once, now disagreeing, with nothing saying which one came later in meaning rather than in time. When a decision supersedes an earlier one here, it points at that earlier one by name and states what it contradicts. The old one is retired with the reason attached, so the history stays auditable by a person, and what an assistant reads is the current answer rather than both.
The people who would notice are the ones who can correct it
Some things go out of date without anyone deciding: the external tool changed, the number moved, the source now says something else. When someone opens that source in the course of their work and finds it contradicts what is recorded, the correction is a proposal that names the entry it replaces, and a person approves it. That keeps two properties that pull against each other — nothing is quietly overwritten by an assistant acting on its own, and nothing wrong stays in force just because correcting it was somebody's side errand.
What it looks like in practice
A team records a decision in March: while the migration is running, no schema change ships without a second reviewer. It is the right call, and every assistant on the project starts repeating it. The migration finishes in June. Nobody deletes the rule, because deleting it is nobody's job and it was never wrong — it was right, and then it quietly stopped applying. Through July, assistants keep telling new engineers to find a second reviewer for changes that no longer need one. The rule is not corrupt or mistaken. It outlived the condition that made it true, and nothing in the record knew that condition existed. Written with its ending attached, the same rule says in March that it dies when the migration completes. In June it stops being served, without anyone having to remember it on the day it mattered.
Questions people ask about this
- Who decides that a memory has expired?
- The condition is written when the memory is, and it has to be concrete enough to check: an explicit revocation, a date, or a named event. When a piece of work makes that condition true, the memory is retired as part of recording that work, with a line saying what made it true. A person can reverse that from the panel — retirement being reversible is what makes it safe to do without asking.
- What about entries that are simply wrong, rather than expired?
- Those take a correction rather than an expiry, and it follows the same path: whoever found the contradiction proposes the replacement, naming the entry it contradicts, and a person approves it. What to avoid is saying it out loud in a conversation and moving on, because then the wrong entry stays in force for everyone else while the conversation ends.
- Doesn't all this bookkeeping stop people writing anything down?
- It is one extra sentence at the moment of writing, and it is the sentence the author is best placed to write: they know whether this is permanent, seasonal or conditional, and they will never know it more clearly than right then. The alternative is not less work. It is the same work done later by someone who has to reconstruct whether a two-year-old line still applies, usually while an assistant is already quoting it.
Where this is verifiable
Product documentation on this site (How it works, Install), the answer on why retrieval does not replace sanctioned memory for the mechanism side of this question, and the expiry, supersession and human-approval rules in the sanctioned product spec. Everything described here is behaviour the tools apply today, not roadmap.
https://www.arroway.app/en/answers/stale-memory-is-worse-than-no-memory