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

TL;DR

MemSkill reframes LLM agent memory as a set of learnable and evolvable memory skills, organized in a shared Skill Bank. A controller selects a Top-$K$ subset of skills per text span, while an executor applies them to build a trace-specific Memory Bank; a fixed designer periodically refines or expands the skill set based on hard-case feedback, forming a closed-loop training loop. Experiments across LoCoMo, LongMemEval, HotpotQA, and ALFWorld show consistent improvements over baselines and strong generalization across base models and datasets. The work provides insights into how adaptive memory behaviors can be discovered and evolved, advancing self-improving memory management for long-horizon tasks.

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 \textbf{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 \emph{controller} that learns to select a small set of relevant skills, paired with an LLM-based \emph{executor} that produces skill-guided memories. Beyond learning skill selection, MemSkill introduces a \emph{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.

Figures (3)

• Figure 1: MemSkill architecture overview. Given an interaction trace, MemSkill processes it span by span: the controller selects a Top-$K$ subset of skills from a shared skill bank conditioned on the current text span and retrieved memories, and an LLM executor applies the selected skills in one pass to update the trace-specific memory bank. The constructed memory is then evaluated on memory-dependent training queries to provide task reward for optimizing the controller, while query-centric failures are logged into a sliding hard-case buffer. Periodically, the designer mines representative hard cases to refine existing skills and propose new ones, yielding alternating phases of skill usage and skill evolution. More skill case study can be found in Section \ref{['sec:case_study']} and Appendix \ref{['appendix:case_study']}.
• Figure 2: Skill generalization under distribution shift on HotpotQA. We transfer the LoCoMo-trained skill bank to HotpotQA and evaluate three context-length settings (50/100/200 concatenated documents) following yu2025memagent. Bars show LLM-judge (L-J) under LLaMA with different Top-$K$ skill counts, compared to MemoryOS and A-MEM.
• Figure 3: Case study. We show representative evolved skills learned on LoCoMo and ALFWorld. ("Description" is omitted for brevity.)