Qwen2.5-7B accuracy on FactConsolidation drops from 45% to 32% as BM25 evidence grows from 64 to 512 tokens

The MemoryAgentBench experiment isolates retrieval from generation on 6K-history traces. With frozen prompts and four-statement evidence, swapping one unrelated statement for an obsolete competing version reduced accuracy from 100% to 44.6% for Qwen2.5-7B and from 100% to 87.8% for Qwen3-14B; all 50 harmful flips were classified as stale. A 32K-history replication reproduced the pattern, with BM25-64 outperforming BM25-512 at 69% versus 58%.

These results demonstrate that an obsolete statement harms performance more than unrelated text of identical length. The failure mode is not missing evidence but failure to suppress a prior version once both versions are co-present in context. This matches patterns observed in earlier RAG studies where distractors in the middle or late positions degrade reasoning even when gold facts remain accessible.

Operationally, agent memory systems require explicit staleness detection or versioned consolidation layers beyond retrieval ranking. Without such mechanisms, scaling context windows or top-k will amplify rather than mitigate knowledge obsolescence errors in deployed agents.

Future benchmarks must separate retrieval hit rate from update application rate and report paired-flip accuracy under controlled obsolete-statement injection.