Home / Vault & Second Brain / A Fresh Date Is Not the Same as the Truth

A Fresh Date Is Not the Same as the Truth

Archive of layered memory cards with a fading timestamp

Last week I wrote two posts about agent memory. First that memory is a structure problem, not a search problem. Then that it is not a storage problem at all, but a behaviour problem. Each time I thought I had reached the bottom of it. I had not. There is a layer underneath that I never questioned, and this week one line from a Reddit thread put its finger on it: a fresh date and the truth are not the same thing.

My agents keep a small block of “world state” — who is home, which systems are up, what happened last night. It gets injected at the top of every session so the agent never starts blind. This week I checked the timestamp on the block my morning session was actually reading. It was three days old. The system had been happily injecting it as if it were current.

Nothing was wrong with the data. It was recorded correctly on Monday and it was accurate on Monday. The failure was that nobody had asked whether it was still accurate on Thursday. The slot had no expiry, so it came up “valid” every time — and valid is not the same as fresh.


Recorded is not the same as verified

The fix that came out of that thread is almost embarrassingly simple, and it is cheaper than any decay algorithm I had been sketching. Stop treating a memory as one thing with one date. Give it two: when it was recorded and when it was last verified. When the underlying source changes, you do not delete the old claim — you mark it stale. Keep it. It is still true that “source X said Y on Tuesday”; what has quietly stopped being true is the idea that Y is current.

That distinction matters more than it sounds. My instinct for stale data has always been to throw it away. But throwing it away destroys the provenance — you lose the ability to see that something changed and when. Marking it stale keeps both: the record and the doubt. The agent gets to know not just what it knows, but how old its certainty is.

The other half of the problem is speed. Some facts change hourly and should never be cached at all — you fetch them at the moment you need them. Others are stable for months and deserve to be stored. Treating every fact the same way is what got me here: a value that should be fetched live, sitting in a cache with no expiry.


Silent wrong output is the expensive kind

The same week, a post on building agents landed the matching principle: build the failure path before the happy path. A demo only has to work. A real agent has to fail correctly. The line that stayed with me was that silent wrong output for two weeks is worse than a loud failure on day one.

Three days of a stale world state is exactly that. Nothing crashed. No error was logged. The agent just quietly reasoned from a Monday version of the world on Thursday. It is the most expensive kind of bug because it never announces itself — you only find it by going back and actually checking a timestamp.

The prescription from that post was three things, and I recognise my own gaps in all of them: a validation step that rejects malformed input instead of guessing, an explicit escape hatch that stops and tags a human rather than improvising, and a log of every decision with the input that caused it — so you can reconstruct what happened at 2am.


What I am measuring next

One metric, because a number beats a feeling: what is my stale rate? How many of the blocks and facts my agents read as current are actually older than their useful life? I have a presence slot that is a proven case. I suspect it is not the only one. And the fix I am going to try is small — inject with an honest marker, (outdated, last checked Monday), instead of presenting it as the present.

A memory system is only as good as the moment it is used. You can store everything perfectly and still hand your agent a slightly old version of the world. The question I never asked was simply: how old is our truth, and who measures that?

If you track the age of what your agent treats as current — or you have never checked and now feel like you should — I would like to hear how you handle it. Comments are open.

Wat vond je van dit bericht?

Leave a Reply

Your email address will not be published. Required fields are marked *