AI agents often repeat the same errors because each run can be treated as a fresh start. Google Research's WikiSkill takes a different approach: it gives an agent a durable place to store what happened, what worked, and what failed, then turns that record into reusable guidance for future tasks.
The result is not continuous learning in the deeper sense. The model itself is not being retrained after every attempt. Instead, WikiSkill builds better instructions around the agent, using a wiki-like memory to shape future behavior.
A wiki for agent experience
WikiSkill is built around a simple problem in AI agent design. If an agent completes a task, hits an error, or discovers a useful strategy, that experience can vanish once the run ends. WikiSkill preserves it.
The framework collects knowledge from both failures and successes. That knowledge is organized in a persistent structure and then packaged into "Agent Skills," which are reusable modules that guide the agent during later work.
This matters because the agent can improve its operating instructions without changing what the model learned during training. In practical terms, the system remembers more about how to act, even though the underlying model does not truly learn continuously.
The work draws on a perspective from Andrej Karpathy about an "LLM Wiki," where experience is compiled into persistent, cumulative knowledge. WikiSkill applies that idea to automatic skill development for AI agents.
The three layers of WikiSkill
WikiSkill separates an agent's workspace into three tiers. Each tier has a different job, which helps the system preserve raw evidence, refine it into knowledge, and test whether new instructions actually help.
- Raw Layer: This stores complete execution traces, including tool calls and results. The data is immutable and acts as the source material for later analysis.
- Wiki Layer: This turns raw traces into structured insights, such as failure patterns and successful strategies. According to the researchers, this layer never resets and grows with each iteration.
- Skill Layer: This contains the active procedural instructions the agent uses while carrying out tasks. Unlike the wiki, these skills can be rolled back if a change reduces performance.
The workflow starts when an inference agent performs tasks using the current skills. Those runs generate execution traces. A "Wiki Maintainer" then reviews the traces, identifies repeated problems and useful tactics, and writes the findings into the wiki.
Next, a "Skill Proposer" uses the updated wiki and execution data to suggest changes to the agent's skills. A gating mechanism tests each proposed change on a separate validation set. If the change helps, it can be kept. If it does not, the skill is rolled back.
The wiki itself remains intact even when a proposed skill fails. That is an important design choice because failed attempts still provide useful information. The system can document what was tried and why it did not work, giving later iterations more context instead of forcing them to start over.
What the tests showed
The researchers tested WikiSkill across five benchmarks: math reasoning, web search, spreadsheet manipulation, document question-answering, and interactive tasks in a virtual environment. The models used were Qwen (4B, 9B, 27B), Gemma-4-31B, and Gemini-3.5-Flash.
In the study, WikiSkill consistently outperforms all previous skill evolution methods. The average gains reported for larger models were substantial. Gemini-3.5-Flash improved from 49.5 percent to 68.1 percent, while Qwen-3.6-27B rose from 39.4 percent to 63.3 percent.
Some benchmark-specific gains were even larger. Gemini-3.5-Flash increased from 33.0 percent to 72.6 percent on LiveMath and from 50.5 percent to 76.6 percent on SpreadSheet.
The improvements were not evenly distributed across every task. Math problems and spreadsheet manipulation saw the biggest gains. Tasks with long document contexts, such as OfficeQA, showed much smaller improvements.
The researchers say smaller models like Qwen-3.5-4B can struggle to reliably execute evolved multi-step search strategies across long contexts. In those cases, the model may fall back to its default behavior instead of using the evolved strategy well.
Why model size still matters
WikiSkill appears to help larger models more. That finding is important because a better skill system still depends on the model's ability to follow and apply the resulting instructions.
At the same time, the framework can help smaller models close part of the performance gap. Smaller models running WikiSkill can match larger models that do not use the framework. That suggests persistent agent memory can be valuable even when the base model is less capable.
The study also found that skills developed by one model can often transfer to another. In some cases, transferred skills worked better than skills the receiving model created by itself. But this was not guaranteed, so transferability should be checked case by case.
This points to a practical tradeoff for AI agent systems. Reusable skills can make agents more effective, but the usefulness of those skills depends on the task, the model, and whether the model can reliably execute the procedure it is given.
A workaround, not true continuous learning
The most important distinction is that WikiSkill does not solve continuous learning for AI models. The model is not permanently updating its internal training after each run. Instead, the surrounding agent system writes, tests, and reuses procedural knowledge.
That makes WikiSkill more like an external memory and instruction improvement loop than a self-learning model. It is less direct than real learning, and the source article notes that this approach is probably more error-prone. Still, the study shows that the workaround can be effective.
For AI agents, that may be enough to matter. If an agent can preserve the lessons of past attempts, avoid repeating documented mistakes, and keep only the skill changes that pass validation, it becomes more useful over time without requiring the model itself to be retrained.