Contradictions are one of the five ways AI memory fails, and the failure runs both ways. A memory that misses them hands your AI conflicting facts. A memory that flags too many teaches you to ignore it.
Why are false alarms worse than they look?
Because people stop listening. A checker that flags every change and every loose phrase as a conflict trains you to dismiss it. Once you're dismissing flags, a real contradiction slips through with the rest. An alarm is only worth having if you believe it when it sounds.
Think of a smoke alarm that goes off every time you make toast. Within a week it's unplugged, and then it can't warn you about a real fire. A contradiction checker works the same way, so Edwin's starts from precision: a flag should mean something.
Isn't every change a contradiction?
No. A budget that moved from $500K to $750K is an update: both facts were true, at different times. Edwin stores every fact with the date it was true, so it can tell an update from a real conflict before it raises anything.
That's why dates come first. As dated facts versus raw transcripts showed, a memory that knows when each fact was true doesn't mistake a change for a conflict. With source dates, Edwin classified 50 of 50 updates correctly; without them, 23 of 50 (internal test, Feb 2026).
When does Edwin count a contradiction?
Only when two different AI providers agree it's a contradiction, or when you confirm it. One model's hunch isn't enough. Agreement between two independent providers means more because disagreement was possible.
How well does the detector do on its own?
Well enough to trust its flags. On an internal test of 150 pairs, Edwin's detector raised zero false alarms while catching 68 of 100 real contradictions (internal test, Apr 2026). That test measured the detector by itself. The rule that two providers must agree came later and now sits on top of it.
Why put precision ahead of catching everything?
Missing a contradiction and inventing one are both errors, but they cost you differently. A miss is a gap. A false alarm is a habit: once you learn to wave flags away, you wave all of them away, real ones included. So Edwin's detector is built to stay quiet unless it has good reason to speak, and a flag still needs a second provider's agreement, or yours, before it counts.
What happens when Edwin finds one?
Edwin flags it, so the conflict is visible to you and your AI instead of buried in your history. The memories involved carry their labels, and Disputed marks a fact the providers disagree about. You stay in charge of what your AI relies on.
What should you ask of any contradiction checker?
- How often does it flag something that isn't a conflict?
- Can it tell an update from an error?
- Who has to agree before a flag counts?
- Has anyone measured it, and can you see the result?
For technical readers: what does the test show?
In standard terms: no false positives on the test set, and recall of 68 of 100. It measures the detector itself; the two-provider rule on top is newer. The test and its context are in our experiment log, and the labels that carry a contradiction to your AI are explained in show your AI what's been checked.