What Should an Agent Remember? Disentangling retention from retrieval in bounded-memory evaluation (arXiv:2610.00366v1)
This arXiv preprint introduces a streaming-recall benchmark that crosses retention and selection rules and evaluates every condition on the same 300 seeded episodes to separate retention (what is stored) from selection (what is surfaced). Key findings: holding access fixed, query-aware selection improves required-fact recall by 15.5 percentage points (95% CI 12.8–18.2); mixed comparisons that also change history access report a 68.7-point advantage of which 53.2 points are attributable to access; under bounded retention, query-aware, dense, and oracle selection hit the retention ceiling and all 319 failures in the bounded-recent condition were due to eviction rather than ranking errors; recall falls to 0% for sufficiently distant targets; on SQuAD dense retrieval can outperform lexical retrieval.