A stricter bar for users’ values and beliefs reduced unsupported saved facts in an AI-agent study

A post sharing a Meta paper says claims about users’ values and beliefs accounted for 21.5% of candidate memories yet contributed more unsupported claims than all other categories combined. Tightening the cutoff only for those claims reduced unsupported saved facts from 6.2% to 4.0% compared with a global threshold saving a similar share, while retaining about 13 points more of the source’s key facts. The results came from synthetic personas.