Last week's post showed why dated facts beat raw transcripts: a date tells your AI when a fact was true. This post is about the other half of trust: whether anyone has checked it.
Why does your AI need to know what's been checked?
Because to your AI, every memory reads the same. A guess from a hurried meeting and a fact that has held up under scrutiny arrive as identical sentences. Without a signal, your AI weighs them equally, and a confident mistake gets the same trust as a checked fact.
People handle this with a habit so natural we rarely notice it. We ask where a claim came from and whether anyone has checked it. A model's own confidence can't stand in for that check, as the first post in this series explains. Confidence describes how sure the model sounds. A check describes what happened when someone tried to prove it wrong.
What does Edwin show beside each memory?
One of four plain labels that say what has actually happened to it: Survived challenge, Not confirmed, Disputed, or Not yet challenged. A second AI provider challenges stored facts in the background, and each memory's label reflects the result.
- Survived challenge: a second AI provider tried to dispute it and couldn't.
- Not confirmed: it was challenged and the support fell short.
- Disputed: the providers disagree about it.
- Not yet challenged: no second provider has checked it yet.
"Not yet challenged" doesn't mean wrong. It means untested, and the label says so plainly. That honesty is the point: your AI can lean on what has survived a challenge and treat the rest with care.
Do the labels actually improve answers?
Yes. Checking facts in the background and showing the answering AI those labels raised accuracy 5.8 points on STALE, a third-party benchmark built on facts that change (pre-registered, Oct 2026). Most of that gain came from the labels themselves: removing them alone took away 4.1 points.
A check at question time helps too: on the same benchmark, it raised accuracy 5.3 points (Oct 2026). Both results point the same way. Give your AI evidence about its memories, not just the memories, and it answers better.
What does a label change in practice?
Say your memory holds two facts about a vendor contract. One has survived a challenge; the other hasn't been challenged yet. Ask your AI about the renewal terms and, with the labels in front of it, it can lead with the checked fact and tell you the other one is untested, instead of blending both into one confident answer. You get the same information, plus the part that matters most: how far to trust each piece.
Why does the check come from a second provider?
A check is only worth having if it can disagree. Different AI providers build their models differently, so a challenge from a second provider is a more independent look than a model reviewing its own work. When a fact survives that, it means something. When the two disagree, you see it: that's what Disputed is for.
What should you ask of any AI memory?
Two questions cut through. Does it tell you what's been checked, separately from how confident it sounds? And can you see evidence that its checks make answers better? Edwin answers the first with a label on every memory and the second with a public measurement record.
For technical readers: what exactly was compared?
The 5.8-point gain compares the background check switched on against switched off, with the labels shown to the answering model. Its 95% confidence interval runs from 2.4 to 9.1 points. Removing the labels alone took away 4.1 of those points, so most of the value reaches the answer through the label. The analysis plan was registered before the run, and the full table is in our October 2026 measurements.