Google Research has released WikiSkill, a new framework designed to give AI agents persistent memory across task executions. Instead of resetting agent experience after each run, WikiSkill logs execution histories in a three-tier workspace—comprising a Raw Layer for execution logs, a Wiki Layer for distilled insights, and a Skill Layer for actionable procedural instructions. While the underlying foundation models do not continuously update their weights, the system allows agents to dynamically refine their own prompt instructions based on past failures and successful strategies.
According to the study, WikiSkill organizes information so that failed instruction changes can be safely rolled back while retaining documented failure reasons within the wiki layer. The researchers evaluated the framework across five task benchmarks using multiple models, including Gemini-3.5-Flash and Qwen-3.6-27B. On average, WikiSkill significantly boosted task accuracy, raising Gemini-3.5-Flash’s performance score from 49.5 percent to 68.1 percent and Qwen-3.6-27B from 39.4 percent to 63.3 percent, with particularly dramatic gains on structured tasks like math and spreadsheet manipulations.
The framework demonstrates a viable, non-parametric workaround for the continuous learning problem in AI agents. While smaller models still struggle with executing complex multi-step search strategies using WikiSkill, larger frontier models achieve massive performance lifts without requiring costly backpropagation or continuous retraining.
Why it matters
Engineers can bypass continuous model retraining by using dynamic, external knowledge bases to significantly improve agent task accuracy.
Enterprise operators can achieve higher reliability on multi-step workflows like spreadsheets and math reasoning using existing inference models.
Source: the-decoder.com
