MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents
Haozhen Zhang, Quanyu Long, Jianzhu Bao, Tao Feng, Weizhi Zhang, Haodong Yue, Wenya Wang
2026-02-06
Summary
This paper introduces a new way for AI agents, powered by Large Language Models (LLMs), to remember things during conversations or while completing tasks. It's about making their memory systems more flexible and adaptable.
What's the problem?
Current AI agent memory systems are limited because they use a fixed set of rules for deciding what to remember and how to update that memory. These rules are designed by humans and don't always work well in different situations or when the AI needs to remember things over a long period of time. Essentially, the memory is too rigid and can't easily adjust to new information or complex tasks.
What's the solution?
The researchers developed a system called MemSkill. Instead of fixed rules, MemSkill treats memory operations like 'skills' that the AI can *learn* and *improve* over time. A 'controller' decides which skills to use, and an LLM then executes those skills to create or update memories. Crucially, there's also a 'designer' that looks at cases where the AI's memory isn't quite right and suggests new or improved skills. This creates a cycle of learning and evolution, making the memory system better and better.
Why it matters?
This work is important because it moves AI agents closer to having truly adaptive and reliable memory. By allowing the memory system to learn and evolve, these agents can handle more complex tasks, remember information more effectively over longer interactions, and ultimately be more helpful and intelligent. It's a step towards AI that can learn from experience and improve its own abilities.
Abstract
Most Large Language Model (LLM) agent memory systems rely on a small set of static, hand-designed operations for extracting memory. These fixed procedures hard-code human priors about what to store and how to revise memory, making them rigid under diverse interaction patterns and inefficient on long histories. To this end, we present MemSkill, which reframes these operations as learnable and evolvable memory skills, structured and reusable routines for extracting, consolidating, and pruning information from interaction traces. Inspired by the design philosophy of agent skills, MemSkill employs a controller that learns to select a small set of relevant skills, paired with an LLM-based executor that produces skill-guided memories. Beyond learning skill selection, MemSkill introduces a designer that periodically reviews hard cases where selected skills yield incorrect or incomplete memories, and evolves the skill set by proposing refinements and new skills. Together, MemSkill forms a closed-loop procedure that improves both the skill-selection policy and the skill set itself. Experiments on LoCoMo, LongMemEval, HotpotQA, and ALFWorld demonstrate that MemSkill improves task performance over strong baselines and generalizes well across settings. Further analyses shed light on how skills evolve, offering insights toward more adaptive, self-evolving memory management for LLM agents.