When LLM Agents start running long-horizon tasks, "can't remember, can't use, can't find accurately" becomes a new infrastructure bottleneck. Shuo Ji et al. at the National University of Singapore released the MRAgent framework on arXiv (2606.06036), accepted at ICML 2026. The core idea: shift memory from a static "retrieve first, then reason" pipeline to a "reason while digging" model.
The framework treats memory operations as first-class actions alongside game actions: reading, writing, searching, appending, and creating files. A meta-reviewer (a strong LLM) reads full game trajectories and iteratively rewrites prompts, file structure, and available operations; in parallel, a memory expert is fine-tuned separately (LoRA) from the agent's own successful trajectories, so the "look up before writing" habit is internalized into parameters. The game policy model's weights stay frozen — memory is layered cleanly on top of existing capability.
On three long-horizon procedurally generated games — Crafter, MiniHack, NetHack — optimizing memory alone raises base-agent scores 2-4×, and a 32B open-source model surpasses Qwen2.5-72B-Instruct (twice the parameter count) on all three games, approaching Claude Opus 4.5 and Gemini 3.1 Pro Thinking. Inefficient-action rate drops 32-65%, repeat-write rate drops 68-83%.
The most worth-noting thing is not another SOTA, but a methodological signal: outside the parameter-and-compute race, "how to manage memory" can be decoupled and optimized independently, with potentially more leverage than stacking parameters. The current engineering bottleneck of Agent systems often isn't the model itself but memory and state management in long-horizon tasks. AutoMem provides a reusable engineering paradigm — treat memory operations as first-class citizens, decouple fine-tuning the memory expert from the policy model, and bring meta-AI into the auto-tuning loop.