Stanford Trains a 32B Model's Memory to Match Opus 4.5

AutoMem lifted a 32B open-weight model to parity with Claude Opus 4.5 and Gemini 3.1 Pro Thinking, using memory training alone, per a new Stanford paper.

That's a 2x to 4x gain across three game-based benchmarks, using the same base model, with no bigger checkpoint required.

The system runs two meta-LLM loops. One reshapes the agent's scaffold: prompts, file schemas, and the vocabulary of actions available to it. The other trains a memory specialist from the agent's own traces, so it learns what to keep and drop from experience.

Filesystem operations count as memory actions, on equal footing with task actions, and the model decides when to use them. The results come from three environments: Crafter, MiniHack, and NetHack.

A 32B model with trained memory matching bigger models makes memory design the cheap lever, not a solved problem. The paper doesn't say how this holds up on tasks with real tool use instead of game state.