Tag
This research demonstrates that continuously updating LLM agent memories through distillation and consolidation loops causes performance regression, even when trained on ground-truth solutions. The study finds that episodic-only retention outperforms text-based consolidation, highlighting significant flaws in current self-improvement paradigms.