Recent research reveals that large language models (LLMs) can produce faulty consolidated memories when continuously updated, leading to a decline in memory utility. This suggests that relying on episodic memories may be more effective for developing robust AI agents than frequent consolidation of experiences.