Researchers at Google Research have introduced WikiSkill, a framework that pairs AI agents with a persistent knowledge base. The idea is simple: instead of throwing away what the agent learned after each run, collect knowledge about failures and successes in a wiki-like structure and use it to improve its abilities over time.

That knowledge gets packaged into "Agent Skills," reusable modules that guide the agent's behaviour without changing what it learned during training.

So the model does not truly learn in a continuous sense — that remains an unsolved problem. It writes better instructions for itself after each run and can pull them up the next time. It is not elegant, and it is probably more error-prone than real learning, but the study shows it is an effective workaround.

The work draws on Andrej Karpathy's idea of an "LLM Wiki": compiling experience into persistent, cumulative knowledge.

Three layers

  • Raw Layer. Stores complete execution traces, from tool calls to results. This data is immutable and serves as raw material.
  • Wiki Layer. Raw data is distilled here into structured insights: documented failure patterns and successful strategies. This layer never resets and only grows with each iteration.
  • Skill Layer. Holds the active procedural instructions the agent follows. Unlike the wiki, skills can be rolled back if an update hurts performance.

The loop runs like this: an inference agent executes tasks with the current skills and generates traces. A "Wiki Maintainer" analyses those traces, spots failure patterns and successful strategies, and writes the findings into the wiki. A "Skill Proposer" uses the updated wiki to suggest targeted skill changes. Finally a gating mechanism tests the proposed change on a separate validation set to confirm it actually helps.

If it does not, the skill is rolled back but the wiki stays intact. Even failed proposals are not lost: the wiki documents what was tried and why it failed, so later iterations build on that.

The results

The framework was tested across five benchmarks: mathematical reasoning, web search, spreadsheet manipulation, document question-answering, and interactive tasks in a virtual environment.

On average, Gemini 3.5 Flash rose from 49.5 to 68.1 percent and Qwen-3.6-27B from 39.4 to 63.3 percent. On individual benchmarks the jumps can be larger: Gemini 3.5 Flash climbed from 33 to 72.6 percent on LiveMath.

Where it does not work

Gains vary a great deal by task type. The largest improvements come in maths and spreadsheets; tasks needing long document contexts show markedly smaller gains.

The researchers' explanation: small models like Qwen-3.5-4B struggle to reliably execute evolved multi-step search strategies across long contexts and fall back to their default behaviour. They read the instruction but cannot carry it to the end.

On the other hand, small models running WikiSkill can match the performance of larger models that do not use the framework. Skills developed by one model often transfer to another and sometimes work better than skills the receiving model built itself — but not always, so transferability should be checked case by case.