WikiSkill co-evolves agent skills with a persistent knowledge base to systematically reuse insights scattered across optimization histories.
WikiSkill is a framework introduced in August 2026 by researchers from Google Research and Virginia Tech that enhances AI agent capabilities by co-evolving executable skills with a persistent knowledge base (wiki).
Key Mechanisms Three-Layer Architecture: The system separates Raw Layer (immutable execution traces), Wiki Layer (structured, compounding knowledge of patterns and failure modes), and Skills Layer (active procedural instructions). Persistent Knowledge: Unlike methods where insights scatter across optimization histories, WikiSkill continuously consolidates experience into the wiki, allowing subsequent skill updates to build on accumulated, deduplicated knowledge. * Evolutionary Loop: An Inference Agent runs rollouts, a Wiki Maintainer updates the wiki, and a Skill Proposer suggests updates based on the wiki, which are filtered by a Gating mechanism using validation performance.
Performance and Impact Superior Results: WikiSkill consistently outperforms state-of-the-art baselines like EvoSkill and SkillOpt, as well as no-skill baselines, across five benchmarks and five models. Model Scaling Complement: Skill evolution complements model scaling; while larger models benefit more, smaller models with WikiSkill can outperform substantially larger models without skills (e.g., Qwen-3.6-27B improved from ~39% to ~63% accuracy). * Transferability: Evolved skills transfer effectively across different models and model families, with skills evolved by one model often outperforming skills the model evolved for itself.
Comparison to Baselines | Feature | WikiSkill | EvoSkill / SkillOpt | | :--- | :--- | :--- | | Knowledge Storage | Persistent Wiki (structured, compounding) | Scattered optimization histories / flat feedback | | Insight Reuse | Systematic reuse across iterations | Limited reuse; insights often lost | | Architecture | Three distinct layers (Raw, Wiki, Skills) | Collapses traces and skills into one | | Performance | Consistently beats baselines | Lower accuracy gains |
The framework confirms that persistent knowledge accumulation is critical for effective skill evolution, enabling agents to systematically refine capabilities over time without losing historical context.
The paper introduces WikiSkill, a framework for treating agent improvement as a cumulative knowledge-acquisition problem rather than a sequence of isolated optimization runs. In many agent settings, useful insights—effective action patterns, failure modes, task decompositions, environmental regularities, and reusable heuristics—are generated during optimization but then discarded or remain trapped in individual trajectories. WikiSkill addresses this by maintaining a persistent knowledge base that stores distilled, structured lessons from prior experience. The system “compiles” raw agent experience into reusable skill artifacts, allowing insights to be retrieved, combined, and applied to new tasks or future versions of the agent.
A central contribution is the co-evolutionary loop between agent skills and the knowledge base. Rather than treating memory as a passive archive, WikiSkill lets the knowledge base actively participate in skill refinement: it can surface relevant prior insights, support the construction of new skill primitives, and be updated when new evidence confirms, refutes, or generalizes existing entries. This design turns scattered optimization histories into an evolving corpus of abstractions, making skill development more systematic and transferable. The key insight is that agent performance can improve not only by updating policies or prompts, but also by preserving and reusing the meta-lessons produced during repeated interaction and feedback.
This matters because iterative agent optimization often suffers from poor sample efficiency and repeated rediscovery of the same heuristics. A persistent, compilable knowledge base can reduce that waste by enabling cross-task and cross-run reuse, improving robustness and composability over time. More broadly, WikiSkill reframes agent evolution as a form of institutional learning: instead of each deployment or training episode starting from scratch, the agent accumulates durable knowledge that can be audited, transferred, and built upon, potentially supporting more scalable long-horizon skill development.