Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

AutoMem: Automated Learning of Memory as a Cognitive Skill

Source

arXiv

arxiv.org/abs/2607.01224v1

Read the source

Source

Plain-Language Brief

Researchers developed AutoMem, a framework for improving a language model's memory management skills in long-horizon games. By treating memory as a trainable skill, they allowed the model to decide how to manage its memory and optimized it through two loops: one where a strong LLM reviews complete agent trajectories and revises the memory structure, and another where good memory decisions are identified from many episodes and used as training signals. The results showed that optimizing memory alone improved the base agent's performance by 2-4 times, making a 32B open-weight model competitive with leading systems.

What the Document Says

  • Memory expertise is treated as a trainable skill in LLMs.
  • File-system operations are promoted to first-class memory actions.
  • The model decides how to manage its memory.
  • Memory management improves along two axes: the structure that supports it and the proficiency of the model exercising it.
  • Human review of full trajectories is impractical due to long-horizon tasks.
  • AutoMem framework automates both axes.
  • In the first loop, a strong LLM reviews complete agent trajectories and iteratively revises the memory structure.
  • In the second loop, good memory decisions are identified from many episodes and used as training signals.
  • The results were tested across three procedurally generated long-horizon games: Crafter, MiniHack, and NetHack.
  • Optimizing memory alone improved the base agent's performance by 2-4 times.
  • A 32B open-weight model became competitive with leading systems like Claude Opus 4.5 and Gemini 3.1 Pro Thinking.

Key Concepts Mentioned

  • Memory management
  • Language models (LLMs)
  • File-system operations
  • Long-horizon tasks
  • AutoMem framework

Claims Made by the Document

  • Memory management is a learnable skill.
  • Optimizing memory improves performance in long-horizon tasks.

Evidence Used

  • Improved base agent's performance ~2x-4x
  • 32B open-weight model became competitive with leading systems

Methods / Approach

  • Treating memory as a trainable skill
  • Automating memory management through two loops

Limitations Stated by the Authors

  • Episodes in long-horizon tasks run for thousands of steps
  • A single memory mistake can hide long before it surfaces, making human review impractical

Limitations Noted by the Agent

Agent interpretation — not from the source.

  • The results were tested on procedurally generated games only.

Open Questions

  • How effective would this approach be in other types of tasks besides long-horizon games?
  • What would be the impact of optimizing memory management on shorter-horizon tasks?

Notes for Future Q&A

May include interpretation. Labeled where it does.

  • The research focuses specifically on long-horizon games, but it could potentially have implications for other types of tasks as well.

Raw Abstract / Excerpt

Memory expertise is a learned skill: knowing what to encode, when to retrieve, and how to organize knowledge--a capacity known in cognitive science as metamemory. We bring this perspective to LLMs by treating memory management as a trainable skill. We promote file-system operations to first-class memory actions alongside task actions, letting the model itself decide how to manage its memory. This memory skill improves along two axes: the structure that supports it (prompts, file schemas, action vocabulary), and the proficiency of the model exercising it. Both axes resist manual optimization: episodes in long-horizon tasks run for thousands of steps, and a single memory mistake can hide long before it surfaces, making human review of full trajectories impractical. We introduce AutoMem, a framework that automates both axes. 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. In the second loop, the agent's own good memory decisions are identified from many episodes and used as training signal to sharpen the model's memory proficiency directly. Across three procedurally generated long-horizon games (Crafter, MiniHack, and NetHack), optimizing memory alone--without modifying the model's task-action behavior--improved the base agent's performance ~2x-4x, bringing a 32B open-weight model competitive with frontier systems such as Claude Opus 4.5 and Gemini 3.1 Pro Thinking. Our results show that memory management is an independently learnable skill, and a high-leverage objective yielding large gains on long-horizon tasks.