AutoMem: Automated Learning of Memory as a Cognitive Skill

Authors: Shengguang Wu, Hao Zhu, Yuhui Zhang, Xiaohan Wang, Serena Yeung-Levy

arXiv ID: 2607.01224

Problem: LLMs lack a learned memory management strategy - they don't know what to encode, when to retrieve, or how to organize knowledge over long-horizon tasks.

Key Methodology:

Promotes file-system operations to first-class "memory actions" alongside task actions, letting the model autonomously manage its own memory via two optimization loops
Loop 1: A strong LLM reviews full agent trajectories and iteratively revises the memory structure (prompts, file schemas, action vocabulary)
Loop 2: Good memory decisions from many episodes are identified and used as training signal to directly sharpen the model's memory proficiency

Key Results:

Optimizing memory alone - without modifying task-action behavior - improved base agent performance ~2x–4x across Crafter, MiniHack, and NetHack
A 32B open-weight model with AutoMem became competitive with frontier systems including Claude Opus 4.5 and Gemini 3.1 Pro Thinking

Applied Context: Builders can treat memory management as a separable, trainable skill rather than hand-crafting prompts or RAG pipelines - AutoMem automates the discovery of good memory strategies and yields massive gains on long-horizon agent tasks with no changes to the underlying model's task policy.