Memento-Skills: Let Agents Design Agents (2026)

Free

Read-Write Reflective Learning over executable skill libraries — agents retrieve, execute, reflect, and rewrite their own skills without retraining the base model; evaluated on HLE and GAIA

FreeFree tier

Inputs: text, codeOutputs: text, code

Type

Open Source

Visit Website GitHub (40 stars)

About Memento-Skills: Let Agents Design Agents (2026)

Memento-Skills is 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, using a memory-based reinforcement learning framework with stateful prompts. Reusable skills are stored as structured markdown files, acting as persistent, evolving memory that encodes both behavior and context. The system introduces a Read-Write Reflective Learning mechanism: in the read phase, a skill router selects relevant skills conditioned on the current state; 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 — all adaptation is realized through the evolution of externalized skills and prompts. Evaluated on the General AI Assistants (GAIA) benchmark and Humanity's Last Exam (HLE), Memento-Skills achieved 26.2% and 116.2% relative improvements in overall accuracy, respectively. Code is available on GitHub.

Key Features

Autonomous construction and adaptation of task-specific agents through experience

Memory-based reinforcement learning with stateful prompts

Reusable skills stored as structured markdown files for persistent, evolving memory

Read-Write Reflective Learning mechanism for continual improvement

Closed-loop design enabling continual learning without updating LLM parameters

Generalist agent that designs agents end-to-end for new tasks

Pros & Cons

Pros

Enables continual learning without retraining the base model
Autonomous agent design reduces need for human engineering
Demonstrated sustained performance improvements on GAIA (26.2%) and HLE (116.2%)
Open-source with available code for reproducibility and extension

Cons

Requires initial set of elementary skills (e.g., Web search, terminal operations) to bootstrap
Performance gains are shown on specific benchmarks; general real-world applicability yet to be fully validated
Relies on structured skill library which may need careful design for complex domains

Best For

Automating complex, multi-step AI assistant tasksBuilding adaptive agents that improve over time without manual retrainingDesigning specialized agents for new tasks from scratchResearch in continual learning and agent self-improvement

FAQ

What is Memento-Skills?

Memento-Skills is a generalist LLM agent system that acts as an agent-designing agent. It autonomously creates and improves task-specific agents through experience, using a memory-based reinforcement learning framework with stateful prompts and reusable skills stored as markdown files.

How does Memento-Skills achieve continual learning?

Memento-Skills uses a Read-Write Reflective Learning mechanism. In the read phase, a skill router selects relevant skills based on the current state. In the write phase, the agent updates and expands its skill library based on new experience. All adaptation is realized through externalized skills and prompts, without updating LLM parameters.

What benchmarks was Memento-Skills evaluated on?

Memento-Skills was evaluated on the General AI Assistants (GAIA) benchmark and Humanity's Last Exam (HLE), achieving 26.2% and 116.2% relative improvements in overall accuracy, respectively.

Is the code for Memento-Skills available?

Yes, the code is publicly available on GitHub, as indicated in the paper.