Tag
This research blog post demonstrates that repeatedly rewriting LLM agent experiences into textual 'lessons' often degrades performance rather than improving it. The author finds that episodic memory retention performs better than abstract consolidation across various benchmarks like ARC-AGI and ALFWorld.