Your AI Didn't Misremember. Its Conclusions Outlived Their Reasons.

TL;DR: When an AI gets a fact wrong after you corrected it, the usual diagnosis is bad memory. The real failure is worse: the correction worked, and it still did not matter. A corrected fact leaves behind every summary, profile and judgment that was built on it — still circulating, still trusted, still phrased as current. Memory failure is non-local. Fixing the source does not reach the conclusions, because at read time nothing connects them. The only moment the dependency is visible is the moment the conclusion is written.

Start with the layer everyone thinks they have solved. When a new fact contradicts an old one about the same entity, good systems now do the honest thing: they do not delete the old fact, they stamp it invalid, record what replaced it and when, and keep the history auditable. That is layer one, and it is real progress. It is also where most memory roadmaps stop — one layer out of three.

Layer two has no contradiction to catch. A conclusion can be false without any new fact ever directly contradicting it, because a premise it quietly depends on died somewhere else. In a recent synthetic benchmark — a preprint, so read the number as a benchmark result, not a field accident rate — every record was true when written. No attack, no hallucination. Agents still published 62 percent of downstream objects whose grounds had already been revoked, in the test case because a signing key was revoked more than twenty sessions away and the revocation notice named only the key. The conclusion and the dead premise shared no entity and no retrieval window, so factual memory had, in principle, no way to find the damage. Correct the root all day: the branches never hear about it.

Layer three is why nobody notices. The standard memory benchmarks mostly test recall — can the system remember — with at most a couple of knowledge updates and questions spanning at most a couple of past sessions. Expiry cannot show up on an exam that never expires anything. Newer work stretches updates into the dozens and cross-session references into the hundreds, and scores with a penalty term: credit for correct answers, minus a weighted charge for reusing facts that should have been forgotten. Score only what an agent remembers and you are paying it to hoard.

What I changed in my own system

I run a personal AI operating system for one person — me — so this is not abstract. My own memory import once carried nearly every record at the highest trust grade, including summaries derived from mixed sources: the summary inherited the strongest trust of anything under it. I downgraded 2,439 records in one night, and wrote about it. What I had not fixed was the structure underneath: a corrected fact in my system was retired by a hand-maintained index of superseded versions, and everything derived from it was found, if at all, by a human remembering to grep. That is not a mechanism. That is a hope.

Four changes followed. First, every derived conclusion now records what it was derived from — the exact sources and versions it stands on — at the moment it is written, because that is the only moment the dependency is local and visible. Second, reading is a checkpoint, not a photocopier: if any upstream source has been invalidated or revised, the conclusion is returned as undecided and sent for review, not quoted with its original confidence. Third, trust is chosen per use, not defaulted globally: for anything that merely must not be missed, one credible source is enough to show it; for anything an agent will act on, every source underneath has to hold — a summary gets the weakest trust of its parents, never the strongest. Fourth, my memory tests now report a forgetting score next to the recall score, separately. One number without the other is not a measurement; it is advertising.

Your AI will keep misremembering in the way that matters most: not inventing facts, but faithfully repeating conclusions whose reasons died in another room. Delete nothing. Stamp what expired, record what everything was built on, and make the reader check the foundations before it repeats the house.

— Jiahui Miao, founder of vertciti, building a personal AI operating system for one.