# Memento-Skills: Let Agents Design Agents

**Authors:** Huichi Zhou, Siyuan Guo, Anjie Liu, Zhongwei Yu, Ziqin Gong, Bowen Zhao, Zhixun Chen, Menglong Zhang, Yihang Chen, Jinsong Li, Runyu Yang, Qiangbin Liu, Xinlei Yu, Jianmin Zhou, Na Wang, Chunyang Sun, Jun Wang  
**arXiv:** [2603.18743](https://arxiv.org/abs/2603.18743)  
**Date:** 2026-03-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2603.18743v1/

## Key points

- Authors Huichi Zhou, Siyuan Guo, Anjie Liu, Zhongwei Yu, Ziqin Gong, Bowen Zhao, Zhixun Chen, Menglong Zhang, Yihang Chen, Jinsong Li, Runyu Yang, Qiangbin Liu, Xinlei Yu, Jianmin Zhou and Na Wang introduce Memento-Skills
- Memento-Skills is a generalist LLM agent system that autonomously constructs task-specific agents
- The system is built on a framework with stateful prompts and reusable skills stored as structured markdown files
- Skills encode behavior and context to enable the agent to retain knowledge across interactions
- Memento-Skills continually improves through the Read--Write Reflective Learning mechanism introduced in Memento 2
- The system allows for continual learning without updating LLM parameters through evolving externalized skills and prompts
- Experimental results show significant improvements in accuracy on benchmarks like General AI Assistants and Humanity's Last Exam
- Memento-Skills represents a significant advancement in AI capabilities
- You can explore the code for this innovative system at https://github.com/Memento-Teams/Memento-Skills.

## Summary

In their paper titled "Memento-Skills: Let Agents Design Agents," authors Huichi Zhou, Siyuan Guo, Anjie Liu, Zhongwei Yu, Ziqin Gong, Bowen Zhao, Zhixun Chen, Menglong Zhang, Yihang Chen, Jinsong Li, Runyu Yang, Qiangbin Liu, Xinlei Yu, Jianmin Zhou and Na Wang introduce a groundbreaking generalist LLM agent system known as Memento-Skills. This system serves as an  by autonomously constructing task-specific agents through . Memento-Skills is built on a  with stateful prompts and reusable skills stored as structured markdown files. These skills encode behavior and context to enable the agent to retain knowledge across interactions. The system starts with elementary skills like Web search and terminal operations and continually improves through the Read--Write Reflective Learning mechanism introduced in Memento 2. The Read phase involves a behavior-trainable skill router selecting the most relevant skill based on the current stateful prompt. In the Write phase, the agent updates and expands its skill library using new experiences. This closed-loop design allows for continual learning without updating LLM parameters as adaptation occurs through evolving externalized skills and prompts. Unlike previous approaches that rely on human-designed agents,Memento-Skills empowers a generalist agent to design agents end-to-end for new tasks. Through iterative skill generation and refinement,the system enhances its capabilities over time.Experimental results on benchmarks like General AI Assistants and Humanity's Last Exam show significant improvements in overall accuracy. Overall,Memento-Skills represents a significant advancement in  and  capabilities. The code for this innovative system is available at https://github.com/Memento-Teams/Memento-Skills for further exploration and development.