AutoMem: Optimizing Agent Memory Alone Lifts a 32B Model to Opus 4.5 Territory

arXiv / HuggingFace Daily Papers·medium signal

Stanford's AutoMem (arXiv 2607.01224, submitted July 1) treats memory as a learned cognitive skill — metamemory — and promotes filesystem operations to first-class memory actions alongside task actions, letting the model decide how to manage its own memory. Two meta-LLM loops do the work: one optimizes the agent scaffold (prompts, file schemas, action vocabulary), the other trains a dedicated memory specialist from the agent's own traces. Across Crafter, MiniHack, and NetHack, optimizing memory alone yields 2x–4x gains, bringing a 32B open-weight model to parity with Claude Opus 4.5 and Gemini 3.1 Pro Thinking. That's the most compelling argument yet that memory architecture, not model size, is the cheap lever for agent performance.