Agents That Learn From Failure: Inside Memento-Skills
Memento-Skills is a self-evolving agent framework where agents learn from failures and rewrite their own skills.
Most agent frameworks treat skills as static. You write them once, load them into context, and hope they work. When they fail, you debug manually or try again with the same broken skill.
Memento-Skills takes a different approach. When a skill fails, the system reflects on why it failed, locates the broken skill, rewrites it, and stores the improved version back into the skill library.
Here's how it works:
The framework runs a continuous Read → Execute → Reflect → Write loop.
Read: Retrieve candidate skills from the local library instead of loading every skill into context.
Execute: Run skills in a local sandbox with actual tool calling - file operations, web search, scripts, external systems.
Reflect: When execution fails, the system records what went wrong, updates the skill's utility score, and attributes the failure to specific skills.
Write: Rewrite broken skills, optimize weak ones, or create new skills when nothing suitable exists.
This isn't about accumulating more skills. It's about building a skill library that improves through task experience.
The system was tested on HLE (Humanity's Last Exam) and GAIA (General AI Assistants) benchmarks. Performance improved over multiple learning rounds as the skill library grew from basic atomic skills into a richer set of learned capabilities.
Built for open-source LLM ecosystems - works with Kimi, MiniMax, GLM, and other OpenAI-compatible endpoints.
Comes with 9 built-in skills (filesystem, web-search, PDF, docx, xlsx, pptx, image analysis, skill-creator, dependency install) that serve as the starting point for the evolving library.
It's 100% open source
Link to Memento-Skills in comments!
The post is fundamentally accurate. Here's the breakdown against the current GitHub repo (Memento-Teams/Memento-Skills, MIT license, ~1.5k stars):
Confirmed accurate:
- The Read → Execute → Reflect → Write loop is described exactly this way — retrieve candidate skills instead of stuffing everything into context, execute via tool calling in a local sandbox, reflect by recording state and updating utility scores when things fail, and write by optimizing weak skills or creating new ones.
- The framework tests on HLE and GAIA, showing performance improving across learning rounds as the skill library grows from atomic skills into more sophisticated learned skills.
- It's designed for open-source LLM ecosystems, explicitly naming Kimi/Moonshot, MiniMax, and GLM/Zhipu, plus OpenAI-compatible endpoints.
- It's MIT licensed (open source).
One correction needed: the post says 9 built-in skills — that's slightly out of date. The current version ships 10 built-in skills: filesystem, web-search, image-analysis, pdf, docx, xlsx, pptx, skill-creator, uv-pip-install, and im-platform (the last one, for Feishu/DingTalk/WeCom/WeChat, was added in v0.2.0). The post's list also omits image-analysis and im-platform while otherwise matching the other 7.
Worth flagging: this is presented as "let agents design agents," but it's not just a lightweight skills library anymore — the current repo (v0.3.8) has grown into a much larger deployable system with a desktop GUI, IM platform bridges (Feishu/DingTalk/WeCom/WeChat), a skill marketplace, and background "dream" consolidation daemons. That's not wrong, just worth knowing the scope is bigger than "agents rewrite skill files."
Small nuance: the tagline is literally "Let Agents Design Agents," which matches the post's opening line — that's not paraphrase, it's the actual repo title.
Everything else in the post (attribution to reflection-on-failure, utility scoring, sandboxed execution) checks out against the source repo and README.