A self-writing skill library lifts Humanity's Last Exam accuracy 116% — with zero weight updates

Notes on Memento-Skills: Let Agents Design Agents (arXiv:2603.18743) — Huichi Zhou, Siyuan Guo, Anji Liu, Zhongwei Yu, Ziqin Gong, Bowen Zhao, Zhixun Chen, Menglong Zhang, Yihang Chen, Jinsong Li, Runyu Yang, Qiang Liu, and 5 more · March 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill evolution

Memento-Skills flips who designs the agent: instead of a human hand-building a task-specific system, a generalist LLM builds and refines its own sub-agents through experience — an agent-designing agent. Capability lives in reusable skills: structured markdown files that serve as external, persistent memory, encoding both the behavior and the context it applies to. The agent starts from a small seed — web search, terminal operations — and grows the library from there. Growth runs on Read-Write Reflective Learning, a two-phase loop from a companion system, Memento 2. A read phase lets a trained skill router pick the most relevant skill for the current stateful prompt; a write phase lets the agent update and expand its own library based on what just happened. None of it touches the base LLM's weights. Every bit of adaptation gets externalized into the skill files and prompts themselves.

On GAIA (General AI Assistants) and Humanity's Last Exam, that closed loop of skill creation, routing, and revision produces sustained accuracy gains as the library accumulates: 26.2% relative improvement in overall accuracy on GAIA, 116.2% on HLE. Both results trace back to the skill library evolving — not to any additional gradient update on the model.

Key numbers

GAIA relative accuracy gain	+26.2%
HLE relative accuracy gain	+116.2%
HLE gain vs. GAIA gain	≈4.4× larger

What this means for skill authors

The design choice worth sitting with: skills here are structured markdown files — the same substrate a lot of skill catalogs, SkillFed included, already treat as the unit of portable capability. When an agent writes its own skill file and folds it back into its library, that file becomes the entire record of what it learned. There's no weight update to fall back on, no fine-tune to audit — the markdown either encodes the behavior and the context correctly, or it doesn't. That raises the bar on what counts as a well-formed skill file. It doesn't lower it.

Self-authored skills are also a different trust problem than human-authored ones. A skill an agent wrote for itself mid-task, that happened to work once, is not the same claim as a skill vetted for reuse across users and contexts. That gap — between "passed this one run" and "safe to share" — is exactly where dual-tool scanning earns its keep. As agent-designed skill libraries get more common, the interesting question for a federated catalog isn't whether a skill worked for the agent that wrote it. It's whether it generalizes past the single experience that produced it.

There's a quieter scaling note too. A behavior-trainable router has to keep picking the right skill as the library grows past a handful of entries — retrieval and selection quality, not skill generation, is the part of this loop most likely to degrade first. That's the same bottleneck any multi-skill agent hits once its library stops fitting in a single prompt, federated or not.

Skills related to this research

Tanstack Integration Learn proven patterns for combining TanStack Query, Router, and Start in full-stack projects. This skill covers setup, data flow coordination, caching strategies, and server-side rendering integration to ensure optimal performance and type safety across your stack.★ 201 Tanstack Router Master TanStack Router with structured guidelines covering type safety, route organization, data loading, and search parameter handling. This skill provides critical and high-priority rules to optimize your React routing architecture and prevent runtime errors.★ 201 Clerk Clerk is a router that directs you to specialized authentication skills based on your framework and use case. Whether you're setting up Clerk in Next.js, React, Vue, Nuxt, Astro, native iOS/Android, or managing organizations and billing, this skill identifies your task and points you to the right resource. Covers CLI operations, custom flows, webhooks, testing, and backend API integration.★ 63 Tanstack Router Tanstack Router delivers file-based routing for React with full type safety, route-level data loading, and search parameter validation. It handles code splitting, preloading strategies, and integrates seamlessly with TanStack Query for efficient data management. Use it to build structured navigation with automatic type inference for params, search values, and loader data.★ 13

Related notes

References

Zhou, H., Guo, S., Liu, A. et al. "Memento-Skills: Let Agents Design Agents." arXiv:2603.18743 (2026).
Wang et al. "Memento 2" (2025) — introduces the Read-Write Reflective Learning mechanism this work builds on.