Memento-Skills is an agent framework enabling agents to autonomously learn, rewrite, and evolve their own skills without retraining the underlying LLM. It targets engineers and researchers seeking continual adaptation and self-improvement in AI agents, offering zero retraining cost by managing skills in an external memory.

How It Works

The core of Memento-Skills is a continual Read-Execute-Reflect-Write loop. When a task is submitted, the agent retrieves or generates a skill, executes it, and then reflects on the outcome. Failures are treated as training signals, prompting the system to update skill utility scores or optimize skill code. This read-write loop allows the agent to progressively expand and refine its capabilities through live interactions, focusing on learning better skills from task experience rather than simply accumulating more tools.

Quick Start & Requirements

For developers, installation involves cloning the repository, setting up a Python virtual environment, and running pip install -e .. A pre-built desktop GUI is available for a no-Python/terminal setup. Key requirements include a Python environment and an LLM API key (e.g., OpenAI, Anthropic, Ollama, or compatible endpoints). A TAVILY_API_KEY is needed for web search capabilities. The project site skills.memento.run offers demos and documentation.

Highlighted Details

• Self-Evolving Agent Framework: Memento-Skills provides a complete stack for orchestration, skill routing, execution, reflection, storage, CLI, and GUI, rather than being a wrapper around existing systems.
• LLM Ecosystem Compatibility: Designed to integrate seamlessly with mainstream open-source LLM platforms like Kimi/Moonshot, MiniMax, GLM/Zhipu, and OpenAI-compatible endpoints.
• Skill Self-Evolution Loop: The system actively learns from failures, revises weak skills, and creates new ones, fostering a dynamic and growing skill library.
• Local-First Deployment: Offers practical deployment surfaces including CLI, a desktop GUI, a Feishu bridge, and a local sandbox for isolated skill execution.
• Benchmark Performance: Demonstrates progressive improvement on challenging benchmarks like HLE (Humanity's Last Exam) and GAIA (General AI Assistants).

Maintenance & Community

The project is part of the broader Memento ecosystem, with a Discord community available for discussion and collaboration. The authors are listed in the provided citation for the associated paper (https://arxiv.org/abs/2603.18743).

Licensing & Compatibility

Memento-Skills is released under the MIT license. This permissive license allows for commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

The framework's core strength lies in its learning and evolution capabilities, particularly its focus on learning from failure. While designed for real-world deployment with multiple surfaces, its emphasis is on agent learning rather than solely on immediate deployment stability compared to systems like OpenClaw.