AI agents that remember things about you are more likely to misremember your opinions than your facts.
Researchers tested a memory system by feeding it details from 100 synthetic personas, generating 4,715 candidate facts and opinions for it to decide whether to store permanently. Assertions about a person's values and beliefs were backed by their source text only 77.9% of the time. Everything else - preferences, habits, biographical details - checked out 96.2% of the time. A single confidence threshold applied to both categories couldn't separate them: loosen it enough to keep solid value assertions and unsupported ones slip through; tighten it enough to block the bad ones and solid facts get thrown out too.
The fix was to give values and beliefs their own, stricter confidence bar instead of one global cutoff. That alone cut unsupported retentions from 6.2% to 4.0%, a modest but consistent one-third reduction, while preserving an estimated 13 percentage points more useful memory than a tough global threshold would have allowed. As AI assistants accumulate persistent memory about users, a wrongly stored opinion is uniquely damaging: it quietly becomes treated as settled fact, shaping future responses with no error message and no obvious way for a user to catch it.
A 36% relative cut in errors won't make headlines, but in agent memory, getting the boring category-by-category rules right matters more than a bigger model that still can't tell what you believe from what you merely said once.