New Microsoft paper recommends for long-running agents, check that a lesson is correct and reusable before putting it into persistent memory.

checking agent memories against the environment before saving them made later tasks more accurate and cheaper, so verification should happen at memory-write time.

A finished agent run is not ground truth. It may contain a wrong assumption, an incomplete procedure, or a fact that becomes stale later.

Their fix is: after each task, a separate memory agent gets read-only access to the environment and checks what is worth keeping before it writes anything into long-term memory.

On CLBench, this setup raised pass rate from 39% to 73%, cut queries from 8.8 to 4.7 per task, and reduced task-agent cost from $3.38 to $1.68.