Reliable agent memory needs a different threshold for every claim type

Models that run persistent memory tend to file weak claims alongside verified facts. Researchers have now quantified the asymmetry: of 4,715 candidate claims tested, only 77.9 percent of values-and-beliefs statements were backed by their source, compared with 96.2 percent for every other category. A single global confidence threshold cannot separate the two — it either admits unsupported value claims or discards well-grounded ones.

The fix is a category-conditional threshold. For categories where inference is reliable — facts, events, attributes — the bar is lowered so more valid claims are kept. For categories prone to hallucination — values, beliefs, subjective judgments — the bar is raised and the model abstains when evidence is thin. The approach was evaluated in a cold-start pipeline deployed across 100 synthetic personas, with repeated held-out evaluation on stored data in each fold.

The stricter threshold on values alone cut the unsupported-retention rate from 6.2 percent to 4.0 percent, a relative reduction of roughly 36 percent, consistent across folds. As a complementary signal, the method preserved about 13 percentage points more coverage than a global threshold operating at a comparable retention level (95 percent confidence interval: 9.8–16.0). In other words, solid facts are not sacrificed to block value-noise.

The core takeaway is that retention reliability depends on claim type, not just model confidence. A category-conditional threshold functions as a simple, effective selective predictor exactly at the memory-write point, without adding complex ranking components or separate filtering stages — a minimal architectural change with measurable impact on memory hygiene.

Caveats apply. The experiment ran on synthetic personas in a controlled environment; production behavior with real users may differ. The work appears as a pre-print on arXiv and has not yet been peer-reviewed, so the figures are as reported by the authors, not independently verified. Still, the direction is clear: when memory writes itself, it helps if it knows when to stay silent.