Agent memory research tests whether stored history is useful, efficient, and reliable

HABIT-Bench reports only 33.61% best macro accuracy for habit induction from weak longitudinal evidence, while Memory Tax warns that structured memory can lose to naive RAG. TrawMem reports role-adaptive retrieval gains, and TTME claims 82.8–84.6% fewer FLOPs than test-time training on five math tasks, though single-seed results and narrow evaluations temper confidence. Replication and performance on deployed, long-horizon agents are the next tests.