Memory-agent benchmarks often merge retrieval and answer generation into a single score, making it hard to tell whether a failure stems from missing an update or from not using an already retrieved update. We investigate this distinction on the single‑hop FactConsolidation task of MemoryAgentBench. With Qwen2.5-7B-Instruct on a 6K‑history setting, BM25 accuracy drops from 45% to 32% when the retrieved evidence budget expands from 64 to 512 tokens, even though a gold‑assisted audit finds the newest statement in every retrieved set. Most errors occur on questions that also contain an obsolete version of the statement. Under a frozen prompt and equal four‑statement evidence size, swapping one unrelated statement for the obsolete competing statement reduces accuracy from 100% to 44.6% for Qwen2.5 and from 100% to 87.8% for Qwen3-14B, with all 50 harmful paired flips classified as stale. A 32K‑history replication similarly favors BM25‑64 over BM25‑512 (69% versus 58%). These findings suggest that retrieving the correct fact alone is insufficient; an obsolete competing statement can be more harmful than an unrelated piece of evidence of the same size.
Review