We found similar results when working on our paper for creating LLM agents with metacognition and explicitly called this out in the paper: https://replicantlife.com