TL;DR of the paper, generated automatically.

Agents can check memory lessons against their original runs before acting.

• Fix: lessons link to their source run and a summary of the outcome.
• Results: the submitter says 3.2 points higher on average than lessons alone.
• Code: on GitHub, linked from the paper.