Rohan Paul

New Meta paper. Most agent memory systems use one confidence cutoff to decide which facts about a user to save, but this paper finds it works better to raise the cutoff only for claims about the user’s values and beliefs. Value and belief claims made up 21.5% of candidate memories but contributed more unsupported claims than all other categories combined. What an agent chooses not to store deserves as much design attention as how it stores and recalls memories, since bad writes become trusted facts later. Saved claims get reused later as settled facts, yet value and belief claims were backed by the source only 77.9% of the time, versus 96.2% for other facts. A global cutoff either lets weak value claims in or throws out good facts elsewhere. A stricter bar on values alone cut unsupported saved facts from 6.2% to 4.0% compared with a global threshold that saved a similar share, while keeping about 13 points more of the source's key facts. Find the fact types your extractor overreaches on and gate only those, but test on your own data, since these results come from synthetic personas.