AutoMem Automates LLM Memory Management Learning

Effective memory management is a learned skill, even for humans, involving decisions about what to encode, when to retrieve, and how to organize knowledge. This concept, known as metamemory, is now being applied to Large Language Models (LLMs) through AutoMem, a novel framework that treats memory management as a trainable cognitive skill. AutoMem empowers LLMs to make their own decisions about managing memory by promoting file-system operations to core memory actions, alongside task-specific actions.The framework operates through two automated loops. The first loop uses a powerful LLM to review agent trajectories and iteratively refine the memory structure, including prompts and file schemas. The second loop identifies successful memory decisions made by the agent across many episodes and uses these as training signals to directly improve the model's memory proficiency. This approach, which optimizes memory independently of task-action behavior, led to a 2x-4x performance improvement for a 32B open-weight model on long-horizon games, making it competitive with frontier systems.