Tag
This paper evaluates eight memory systems for LLM agents across five diverse scenarios, finding that giving agents active control over storage and retrieval (rather than passive pipelines) yields the best cross-scenario generalization, leading to the proposed AutoMEM framework.