Memento-Skills: Let Agents Design Agents
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
2026-03-20
Summary
This paper introduces Memento-Skills, a new system where an AI agent can actually *design* other AI agents to complete tasks. It's like an AI that builds specialized AI tools for itself, and it gets better at building them over time without needing to be reprogrammed.
What's the problem?
Traditionally, creating AI agents for specific jobs requires humans to carefully design and program them. This is time-consuming and doesn't allow the AI to adapt easily to new, unforeseen tasks. Existing AI systems often struggle to learn continuously and apply past experiences to new situations without needing constant updates to their core programming.
What's the solution?
The researchers created a system that uses a 'memory' of reusable skills, stored as simple text files. The main agent learns by trying things out, and when it finds something useful, it saves it as a skill. When faced with a new task, it picks the best skills from its memory to use. It then improves those skills based on the results, essentially learning from experience. This happens in a cycle of 'Read, Write, Reflect' – reading relevant skills, writing updates based on what happened, and reflecting on how to improve. Importantly, the core AI doesn't change; it just gets better at using and creating these external skills.
Why it matters?
This is a big step towards more adaptable and generally intelligent AI. Instead of needing a human to build a new agent for every problem, this system can create them automatically. The significant improvements in performance on challenging AI benchmarks show that this approach is effective and could lead to AI that can handle a wider range of tasks with less human intervention.
Abstract
We introduce Memento-Skills, a generalist, continually-learnable LLM agent system that functions as an agent-designing agent: it autonomously constructs, adapts, and improves task-specific agents through experience. The system is built on a memory-based reinforcement learning framework with stateful prompts, where reusable skills (stored as structured markdown files) serve as persistent, evolving memory. These skills encode both behaviour and context, enabling the agent to carry forward knowledge across interactions. Starting from simple elementary skills (like Web search and terminal operations), the agent continually improves via the Read--Write Reflective Learning mechanism introduced in Memento~2~wang2025memento2. In the read phase, a behaviour-trainable skill router selects the most relevant skill conditioned on the current stateful prompt; in the write phase, the agent updates and expands its skill library based on new experience. This closed-loop design enables continual learning without updating LLM parameters, as all adaptation is realised through the evolution of externalised skills and prompts. Unlike prior approaches that rely on human-designed agents, Memento-Skills enables a generalist agent to design agents end-to-end for new tasks. Through iterative skill generation and refinement, the system progressively improves its own capabilities. Experiments on the General AI Assistants benchmark and Humanity's Last Exam demonstrate sustained gains, achieving 26.2\% and 116.2\% relative improvements in overall accuracy, respectively. Code is available at https://github.com/Memento-Teams/Memento-Skills.