Agent Memory benchmarks should measure task success, not retrieval alone