AutoMem treats memory management as a trainable capability for LLM agents. It makes file-system operations first-class actions, then uses two automated loops: one revises the prompts, schemas, and action vocabulary that structure memory; the other extracts good memory decisions from many episodes as training signals. On Crafter, MiniHack, and NetHack, memory-only optimization reportedly improved performance by roughly 2x–4x, making a 32B open-weight model competitive with systems including Claude Opus 4.5 and Gemini 3.1 Pro Thinking. The paper frames metamemory as an independent, high-leverage objective for long-horizon agents.
No heat snapshots are available in the last 24 hours.