Paper

{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.18743v1</id>\n    <title>Memento-Skills: Let Agents Design Agents</title>\n    <updated>2026-03-19T10:45:22Z</updated>\n    <link href='https://arxiv.org/abs/2603.18743v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.18743v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>We introduce \\emph{Memento-Skills}, a generalist, continually-learnable LLM agent system that functions as an \\emph{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 \\emph{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.\n  Starting from simple elementary skills (like Web search and terminal operations), the agent continually improves via the \\emph{Read--Write Reflective Learning} mechanism introduced in \\emph{Memento~2}~\\cite{wang2025memento2}. In the \\emph{read} phase, a behaviour-trainable skill router selects the most relevant skill conditioned on the current stateful prompt; in the \\emph{write} phase, the agent updates and expands its skill library based on new experience. This closed-loop design enables \\emph{continual learning without updating LLM parameters}, as all adaptation is realised through the evolution of externalised skills and prompts.\n  Unlike prior approaches that rely on human-designed agents, Memento-Skills enables a generalist agent to \\emph{design agents end-to-end} for new tasks. Through iterative skill generation and refinement, the system progressively improves its own capabilities. Experiments on the \\emph{General AI Assistants} benchmark and \\emph{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.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.LG'/>\n    <published>2026-03-19T10:45:22Z</published>\n    <arxiv:comment>Memento-Skills Technical Report</arxiv:comment>\n    <arxiv:primary_category term='cs.AI'/>\n    <author>\n      <name>Huichi Zhou</name>\n    </author>\n    <author>\n      <name>Siyuan Guo</name>\n    </author>\n    <author>\n      <name>Anjie Liu</name>\n    </author>\n    <author>\n      <name>Zhongwei Yu</name>\n    </author>\n    <author>\n      <name>Ziqin Gong</name>\n    </author>\n    <author>\n      <name>Bowen Zhao</name>\n    </author>\n    <author>\n      <name>Zhixun Chen</name>\n    </author>\n    <author>\n      <name>Menglong Zhang</name>\n    </author>\n    <author>\n      <name>Yihang Chen</name>\n    </author>\n    <author>\n      <name>Jinsong Li</name>\n    </author>\n    <author>\n      <name>Runyu Yang</name>\n    </author>\n    <author>\n      <name>Qiangbin Liu</name>\n    </author>\n    <author>\n      <name>Xinlei Yu</name>\n    </author>\n    <author>\n      <name>Jianmin Zhou</name>\n    </author>\n    <author>\n      <name>Na Wang</name>\n    </author>\n    <author>\n      <name>Chunyang Sun</name>\n    </author>\n    <author>\n      <name>Jun Wang</name>\n    </author>\n  </entry>"
}