memskill-evolving-memory-operations

name	memskill-evolving-memory-operations
title	MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2602.02474
keywords	["Agent Memory","Skill Learning","Self-Improvement","LLM Controller","Reinforcement Learning"]
description	Build agents that evolve their own memory operations by learning a skill bank of memory transformations and periodically discovering new skills from challenging cases, enabling adaptive memory management that improves with scale.

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

Problem Context

Existing LLM agent memory systems rely on static, hand-designed operations (insert, update, delete) that hard-code human assumptions. These fixed procedures fail when interaction patterns vary, histories grow longer, or task structure changes. Agents cannot adapt their memory strategies or learn what truly matters for their specific workflow.

Core Concept

MemSkill introduces [learnable skill banks, controller-executor architecture, LLM-guided evolution] to enable agents to discover and refine memory operations autonomously. A learned policy (controller) selects which skills to apply per context, while an LLM-based designer periodically identifies missing behaviors and proposes new skills through mining hard cases.

Architecture Overview

• Skill bank: Learnable memory operations (Insert, Update, Delete, Skip) stored as prompts; initially hand-designed
• Controller: Neural policy selecting relevant skills for each context span (trained with PPO)
• Executor: LLM-based component applying selected skills to generate memories
• Designer: LLM mining hard cases to identify skill gaps and propose new operations
• Training loop: Alternate between RL-based skill selection and skill bank evolution

Implementation

Step 1: Define initial skill bank

Create a basic set of learnable memory operations. Each skill is represented as a prompt specifying when and how to transform interaction traces.

# Initial skill bank definition
class MemorySkillBank:
    def __init__(self):
        self.skills = {
            'insert': {
                'description': 'Add new information to memory',
                'prompt': 'Given the interaction: {interaction}, '
                          'extract and add the most important fact to memory.',
                'parameters': {}
            },
            'update': {
                'description': ,
                : 
                          
                          ,
                : {}
            },
            : {
                : ,
                : 
                          
                          ,
                : {}
            },
            : {
                : ,
                : ,
                : {}
            }
        }
        .skill_names = (.skills.keys())

     ():
        
        skill = .skills[skill_name]
         skill[].(
            interaction=interaction, memory=memory
        )

     ():
        
        .skills[skill_name] = {
            : description,
            : prompt,
            : {}
        }
        .skill_names.append(skill_name)