that is the approach also taken in this paper for building LLM agents with metacognition: https://replicantlife.com/