Research project · arXiv:2603.18743
Memento-Skills
Let agents design agents. A self-evolving agent framework built around
skills as first-class units of capability — a library the agent can retrieve,
execute, evaluate, repair, and rewrite over time.
What is Memento-Skills?
Instead of treating tools as a flat list, Memento-Skills treats them as a library. The core loop is Read → Execute → Reflect → Write. When a task fails, the system does not just retry: it tries to identify the weak skill, improve it, and write the improved capability back into memory.
The three paradigms of LLM adaptation. Pre-training and fine-tuning update model parameters and demand large data and compute budgets. Deployment-time learning keeps the model frozen while accumulating experience in external skill memory, enabling continual adaptation from live interactions.
The architecture of the self-evolving agent. The system retrieves or generates skills, executes them, reflects on the outcome, and writes improvements back into the skill library.
Learning results
We evaluate Memento-Skills on two benchmarks: HLE, which probes expert-level reasoning with extremely difficult questions across disciplines, and GAIA, which evaluates general-purpose assistants on real-world, multi-step tasks involving tools, files, and the web.
Performance improves over multiple learning rounds while the skill library grows from atomic tools into a richer, more repairable and more retrievable memory of capabilities.
Core question. Memento-Skills is not centred on how to make an assistant run. It is centred on how to make an agent learn from deployment experience, reflect on failure, and rewrite its own skill code and prompts.
Learn from failure
Failures are treated as training signals, not just reasons to retry.
Rewrite its own skills
The system can optimise prompts, modify skill code, and create new skills when needed.
Run in the real world
Local execution, persistent state, CLI, GUI, and IM integration make it deployable beyond a paper demo.
Why it matters
Not just tool use
The system is designed to improve weak skills instead of only accumulating more tools.
Deployment-time learning
It learns from live tasks while keeping the underlying model frozen.
Practical surfaces
CLI, GUI, Feishu integration, and local sandbox execution make it usable beyond a paper demo.
Deployment surfaces
CLI, desktop GUI, Feishu bridge, skill verification, and local sandbox execution make the system practical for real deployment.
CLI and desktop GUI
memento agent and memento-gui cover both terminal-first and visual workflows.
Feishu and verification
memento feishu and memento verify extend the system into messaging and skill auditing.
One repo, one learning agent
python -m venv .venv && source .venv/bin/activate && pip install -e . && memento doctor && memento agentOne-click GUI install
Prefer a pre-built desktop app with no Python or terminal? That ships as Memento Science — the release page detects your platform and serves the current macOS or Windows build.
Quick start (developer)
git clone https://github.com/Memento-Teams/Memento-Skills.git
cd Memento-Skills
python -m venv .venv
source .venv/bin/activate
pip install -e .
On first launch, ~/memento_s/config.json is created automatically. Fill in your model
profile, then start the app:
memento doctor
memento agent
memento-guiCitation
@article{zhou2026mementoskills,
title={Memento-Skills: Let Agents Design Agents},
author={Zhou, Huichi and Guo, Siyuan and Liu, Anjie and Yu, Zhongwei and Gong, Ziqin
and Zhao, Bowen and Chen, Zhixun and Zhang, Menglong and Chen, Yihang
and Li, Jinsong and Yang, Runyu and Liu, Qiangbin and Yu, Xinlei
and Zhou, Jianmin and Wang, Na and Sun, Chunyang and Wang, Jun},
journal={arXiv preprint arXiv:2603.18743},
year={2026},
url={https://arxiv.org/abs/2603.18743}
}