reflexion

Definition

After a failed attempt, the agent writes a natural-language critique of what went wrong and carries that text into the next attempt as episodic memory. Because the update is verbal rather than gradient-based, the model stays frozen and improvement comes purely from accumulated self-criticism across trials. It is framed as reinforcement learning with language standing in for the reward signal and the policy update.