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

MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 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.

Proposes MemSkill, which learns a controller to select top-K memory skills from an evolving skill bank, an LLM executor that extracts memory with the selected skills, and an LLM designer that refines/adds skills from hard cases, trained with PPO.

Problem

LLM agent memory systems rely on static, hand-designed extraction/consolidation/pruning operations that are rigid across interaction patterns and inefficient on long histories.

Approach

Reframe memory operations as a shared, learnable and evolvable skill bank: an RL-trained controller selects top-K skills per text span, an LLM executor applies them in one pass, and an LLM designer periodically evolves the skill set from hard cases in a closed loop with rollback.

Key results

• On LoCoMo with LLaMA-3.3-70B, MemSkill reaches 53.82 L-J vs 49.71 for the best baseline A-MEM... with 57.36 average L-J across LoCoMo and LongMemEval, the top average among compared methods.
• On ALFWorld with LLaMA-3.3-70B, MemSkill achieves 80.36 average success rate, outperforming LightMem (74.83) by 5.53 points.
• Ablation on LoCoMo (LLaMA): removing the designer drops L-J from 53.82 to 46.50; random skill selection drops it to 48.43.

Surprising findings

• Skills learned on LoCoMo with LLaMA transfer to Qwen and LongMemEval without retraining and still achieve the best results there.
• Without in-context demonstrations, the No-Memory baseline on ALFWorld is far lower, and MemSkill's gains are larger in that controlled setting (57.71 vs 45.99 for LightMem).

Limitations

• Relies on large API-served LLMs (70B/80B) for executor and designer, with costs not fully captured in inference-time cost analysis.
• Performance is sensitive to span size; larger spans (1024, 2048) reduce quality.
• Evaluation is on a limited set of benchmarks with an LLM judge.

Experimental setup

Datasets: LoCoMo LongMemEval HotpotQA ALFWorld AppWorld

Benchmarks: LoCoMo LongMemEval ALFWorld HotpotQA AppWorld

Metrics: F1 LLM-judge accuracy (L-J) success rate steps pass rate

Baselines: No-Memory CoN ReadAgent MemoryBank A-MEM Mem0 LangMem MemoryOS LightMem AWM Expel

Builds on: PPO Gumbel-Top-K Mem0 MemoryOS

Closest papers by meaning

• MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.94)
• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.92)
• MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory (NeurIPS 2026 · similarity 0.92)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.92)
• AUTOMEM: Automated Learning of Memory as a Cognitive Skill (NeurIPS 2026 · similarity 0.91)
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.91)