Announcement_7

Continuously consolidating useful experiences into LLM-written memory can eventually make agents worse than using no memory; preserving raw episodes and gating consolidation is more reliable—see our new preprint Useful Memories Become Faulty When Continuously Updated by LLMs.