WikiSkill Turns Agent Experience into Persistent Knowledge for Skill Evolution

Introduction

An agent that interacts with the world repeatedly needs more than a capable base model or a large context window. It must preserve useful lessons from execution: which steps worked, why an attempt failed, and when a particular workflow should be applied. WikiSkill addresses this problem by proposing a framework in which agent skills and a persistent knowledge base evolve together.

A three-layer view of learning

Recent systems can discover or refine skills from execution experience. Yet the reasoning behind those updates often remains scattered across trajectories, feedback, and optimization records. Valuable observations may influence one iteration without becoming reusable knowledge for the next one. This makes long-term improvement less systematic and can force the agent to repeat earlier exploration.

WikiSkill separates the process into three related layers:

• Raw execution experience: task traces, outcomes, and feedback that provide evidence about agent behavior.
• A persistent wiki: a continuously updated store that consolidates experience into reusable knowledge, patterns, and conditions of use.
• Executable skills: workflows, instructions, or strategies that an agent can invoke during task execution.

The central idea is that experience should not only produce the current skill revision. It should also be compiled into the wiki, allowing future skill updates to build on accumulated knowledge rather than starting from an incomplete history each time.

Main findings

Across diverse models and benchmarks, the paper reports that WikiSkill consistently outperforms existing skill-evolution approaches and improves on no-skill baselines in most model–benchmark settings. The results also suggest that skill evolution complements model scaling. Larger models generally benefit more from evolved skills, but skills can let smaller models outperform substantially larger models that receive no comparable skill support.

The framework also shows promising transfer behavior. Evolved skills can move across models and model families instead of remaining tied to the system that created them. In some cases, a skill evolved by another model performs better than a model’s own self-evolved skill. This points toward a future in which skills are treated as reusable assets that can be shared independently of model weights.

Ablation studies reinforce the role of the knowledge layer. Removing or weakening persistent knowledge accumulation reduces the effectiveness of skill evolution, suggesting that the wiki is not merely a storage convenience. It provides the continuity that connects separate rounds of experimentation.

Why it matters

WikiSkill offers a useful division of labor: experience supplies evidence, the wiki organizes and compresses that evidence, and skills turn it into action. This gives developers a clearer architecture for long-running agents and a possible interface for sharing capabilities across different models.

The available material does not provide detailed benchmark settings, model configurations, or numerical gains, so the findings should be read as evidence for a promising framework rather than a universal guarantee. Practical deployments will still need safeguards against outdated knowledge, accumulated mistakes, and skills being applied outside their valid scope. Even so, WikiSkill highlights an important direction for agent learning: sustainable improvement requires not only better skills, but also infrastructure that preserves and refines the experience from which those skills are built.

Source: Hugging Face Daily Papers