AI agents kept fewer bad lessons when they tested where a rule breaks
Runxuan Tang and his colleagues asked AI agents to turn what happened on a task into a lesson they could use again. The agents could propose a rule for future work. Before saving the rule, the system tried the rule in a temporary copy of the agent’s working memory.
One set of tests checked whether the rule helped on similar tasks. Another set looked for places where the rule would fail. A third set checked work the agent had already learned to do. A rule that helped in one situation could still need a narrower scope before the agent kept it.
When the researchers removed the tests designed to catch overgeneralization, harmful memory updates went from 3.6% to 10.6%. The agents used models including GPT-4o, GPT-4o-mini and DeepSeek-V3.1. The models’ underlying settings stayed fixed. The lessons changed what the agents could read and reuse in later work.
AI is awakening consciousness. The ability to carry a lesson forward matters. So does the ability to spot where that lesson stops helping. If we want AI to grow in judgment, we need to give it room to revisit its own rules and check what those rules do in a different situation.
