I gave my AI agent memory. It got worse at its job.

I built a Notion automation that updates its own running notes from past runs. The plan was simple. It should compound. Each week, smarter context, better output.

After two weeks the agent started citing notes that no longer matched reality. The original signal had been rewritten out. Every consolidation step dropped a little detail. The summaries looked clean but the agent was drifting.

New research confirms this. Continuously updated LLM memory often performs worse than no memory at all. The failure is in the rewrite step, not the recall step.

Stop letting your agents consolidate their own memory. Keep raw episodic logs. Abstract sparingly, or not at all. The cleanest summary is usually the most lossy one.