Research map / Multi-agent systems and games / LLM agents

AUTOMEM: Automated Learning of Memory as a Cognitive Skill

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Treats memory management as a trainable skill: file-system memory actions in the agent's action space, optimized via a meta-LLM scaffold-revision loop and a LoRA-trained memory specialist, lifting a 32B model to frontier-level progression on Crafter, MiniHack, and NetHack.

Problem

LLM agents on long-horizon tasks exceed the context window, and memory management (what to encode, retrieve, organize) is hard to optimize manually because consequences of memory mistakes surface only many steps later.

Approach

Promote file-system operations to first-class actions alongside task actions, then automate two optimization loops driven by a meta-LLM: (1) iteratively revise the memory scaffold (code, prompts, file schema) using full trajectory review, and (2) curate the agent's own good memory decisions as SFT data to LoRA-train a dedicated memory specialist that is deployed alongside an unmodified gameplay mo

Key results

• Scaffold optimization alone raised Qwen2.5-32B-Instruct progression from 25.0% to 47.27% on Crafter, 7.5% to 27.5% on MiniHack, and 0.42% to 1.57% on NetHack, with weights untouched.
• Adding the trained memory specialist (LoRA) reached 51.36% Crafter, 30.0% MiniHack, and 1.85% NetHack, comparable to Claude-Opus-4.5 (49.5/27.5/2.0).
• The trained specialist cut the memory writes-per-SEARCH ratio by 54-72% across environments, indicating more consult-before-write behavior.

Surprising findings

• Optimizing memory alone (without changing task-action weights) also reduced unproductive gameplay actions by 32-65%, suggesting memory and task behavior are coupled in the shared action space.
• A 32B model with optimized memory outperforms Qwen2.5-72B-Instruct under the same harness by a wide margin on all three games.

Limitations

• Memory is episodic; the file system resets each episode, so cross-episode persistent memory is not studied.
• Evaluated only on three procedurally generated game environments; transfer to real-world tasks is untested.
• A separate scaffold and memory specialist is optimized per environment; sharing across environments is unexplored.

Experimental setup

Datasets: BALROG environments (Crafter, MiniHack, NetHack) self-play traces

Benchmarks: BALROG Crafter BALROG MiniHack BALROG NetHack

Metrics: progression rate (%) unproductive action rate empty-search rate redundant write rate input context size

Baselines: Gemini-3-Pro Gemini-3.1-Pro-Thinking Claude-Opus-4.5 Gemini-2.5-Pro DeepSeek-R1 Qwen2.5-72B-Instruct Qwen2.5-7B-Instruct sliding window chain-of-thought

Builds on: BALROG MemAct MemGPT ADAS ExpeL LoRA

Closest papers by meaning

• MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.91)
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.90)
• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.90)
• A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.90)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.89)
• Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents (NeurIPS 2026 · similarity 0.89)