AutoMem treats LLM memory management as a trainable skill, allowing the model to use file-system operations (read, write, search) as first-class actions. A meta-LLM autonomously optimizes memory in two loops: first by revising the memory structure (prompts, schemas) from trajectory reviews, second by training a dedicated memory model on good memory decisions.
On three long-horizon games (Crafter, MiniHack, NetHack), optimizing memory alone improved a 32B model's performance by roughly 2-4x, making it competitive with frontier systems like Claude and Gemini. This shows that memory management is a separable skill that can be automated for large gains, without modifying the model's task actions.