An agent's memory has two separate decisions: what to keep as information arrives (retention), and what to surface when a query appears (retrieval). The authors build a streaming that crosses five retention rules with five selection rules on 300 seeded episodes, varying only one token to make queries informative or not.
When the amount of history each method can access is held fixed, query-aware selection improves recall by 15.5 percentage points. A comparison that also gives some methods full history access reports a 68.7-point advantage, but 53.2 of those points come from access, not query awareness. Under bounded retention, strong selection rules all hit the same ceiling set by retention, and in the bounded recency condition every observed failure is an eviction, not a ranking error. The ceiling persists on SQuAD, but dense retrieval beats lexical on real text, unlike in the synthetic data.
The finding matters because common memory evaluations can mistake a history-access advantage for a better memory algorithm. The paper's conclusion: bounded-memory evaluations should hold access fixed and report retention and selection separately.