K² · Artificial intelligence
Teaching an AI to take better notes made it 2-4x smarter at long games
The authors built AutoMem, a framework that treats memory management in LLMs as a trainable cognitive skill and automates its improvement.
SW
HZ
Shengguang Wu, Hao Zhu, Yuhui Zhang et al.
5 authors · cs.AI, cs.CL, cs.MA
The 30-second scan
A mid-sized AI, given no upgrade to its actual gameplay, jumped 2x to 4x in performance.
- AutoMem promotes file-system operations to first-class memory actions alongside task actions, letting the model itself decide how to manage its memory.
- The framework improves memory along two axes: the structure supporting it (prompts, file schemas, action vocabulary) and the proficiency of the model exercising it.
- In the first loop, a strong LLM reviews complete agent trajectories and iteratively revises the memory structure that shapes how the agent interacts with its memory files.
Optimizing memory alone improved the base agent's performance ~2x-4x on long-horizon tasksA 32B open-weight model was made competitive with frontier systems such as Claude Opus 4.5 and Gemini 3.1 Pro ThinkingEvaluated across three procedurally generated long-horizon games (Crafter, MiniHack, and NetHack)