Google Research has developed WikiSkill, a framework that enables AI agents to retain and reuse knowledge from previous tasks, enhancing their ability to learn from mistakes and successes. The system stores insights about both failures and successes in a wiki-like structure, allowing agents to refine their strategies over time. This approach helps improve task execution without requiring continuous retraining, offering a practical solution to the challenge of persistent learning in AI agents. According to the researchers, while the method is not as elegant as true learning, it serves as an effective workaround for improving agent performance. The framework organizes an agent's workspace into three distinct layers: the Raw Layer for execution traces, the Wiki Layer for structured insights, and the Skill Layer for procedural instructions. These layers work together to refine the agent's behavior through iterative learning and feedback loops. The system includes a Wiki Maintainer that analyzes execution data, identifies patterns, and updates the wiki, while a Skill Proposer suggests targeted skill changes that are tested for effectiveness. If a proposed change fails, the skill can be rolled back, but the wiki remains intact, preserving the knowledge gained from failed attempts. This continuous feedback loop allows the wiki to grow over time, while skills can be adjusted based on performance outcomes. The framework was tested across five benchmarks, including math reasoning, web search, and spreadsheet manipulation, with varying results depending on the task type and model size. The researchers noted that while smaller models benefit from WikiSkill, they still struggle with complex multi-step tasks compared to larger models. However, skills developed by one model can sometimes transfer to another, offering potential for cross-model learning and improvement. The results show that WikiSkill outperforms existing skill evolution methods, with significant gains in performance for models like Gemini-3.5-Flash and Qwen-3.6-27B. The framework's ability to retain and build upon past experiences represents a step forward in making AI agents more adaptable and efficient. The researchers emphasize that while the method is not a complete solution to continuous learning, it provides a valuable tool for improving agent performance through structured knowledge retention and refinement.

The results suggest that WikiSkill could be a useful addition to AI development, particularly for applications requiring persistent learning and adaptation over time. The framework's success in improving performance across various benchmarks highlights its potential for broader implementation in AI systems. The researchers conclude that while the method has limitations, it offers a practical and effective way to enhance AI agent capabilities through structured knowledge retention and iterative improvement. The results indicate that WikiSkill could be a valuable tool for developers looking to improve the adaptability and efficiency of AI agents in real-world applications. The framework's ability to retain and build upon past experiences represents a significant step forward in making AI agents more adaptable and efficient. The researchers conclude that while the method is not a complete solution to continuous learning, it provides a valuable tool for improving agent performance through structured knowledge retention and refinement. The results suggest that WikiSkill could be a useful addition to AI development, particularly for applications requiring persistent learning and adaptation over time. The framework's success in improving performance across various benchmarks highlights its potential for broader implementation in AI systems. The researchers emphasize that while the method has limitations, it offers a practical and effective way to enhance AI agent capabilities through structured knowledge retention and iterative improvement. The results indicate that WikiSkill could be a valuable tool for developers looking to improve the adaptability and efficiency of AI agents in real-world applications. The framework's ability to retain and build upon past experiences represents a significant step forward in making AI agents more adaptable and efficient. The researchers conclude that while the method is not a complete solution to continuous learning, it provides a valuable tool for improving agent performance through structured knowledge retention and refinement. The results suggest that WikiSkill could be a useful addition to AI development, particularly for applications requiring persistent learning and adaptation over time. The framework's success in improving performance across various benchmarks highlights its potential for broader implementation in AI systems.

The researchers emphasize that while the method has limitations, it offers a practical and effective way to enhance AI agent capabilities through structured knowledge retention and iterative improvement. The results indicate that WikiSkill could be a valuable tool for developers looking to improve the adaptability and efficiency of AI agents in real-world applications. The framework's ability to retain and build upon past experiences represents a significant step forward in making AI agents more adaptable and efficient. The researchers conclude that while the method is not a complete solution to continuous learning, it provides a valuable tool for improving agent performance through structured knowledge retention and refinement. The results suggest that WikiSkill could be a useful addition to AI development, particularly for applications requiring persistent learning and adaptation over time. The framework's success in improving performance across various benchmarks highlights its potential for broader implementation in AI systems. The researchers emphasize that while the method has limitations, it offers a practical and effective way to enhance AI agent capabilities through structured knowledge retention and iterative improvement. The results indicate that WikiSkill could be a valuable tool for developers looking to improve the adaptability and efficiency of AI agents in real-world applications. The framework's ability to retain and build upon past experiences represents a significant step forward in making AI agents more adaptable and efficient. The researchers conclude that while the method is not a complete solution to continuous learning, it provides a valuable tool for improving agent performance through structured knowledge retention and refinement. The results suggest that WikiSkill could be a useful addition to AI development, particularly for applications requiring persistent learning and adaptation over time. The framework's success in improving performance across various benchmarks highlights its potential for broader implementation in AI systems. The researchers emphasize that while the method has limitations, it offers a practical and effective way to enhance AI agent capabilities through structured knowledge retention and iterative improvement. The results indicate that WikiSkill could be a valuable tool for developers looking to improve the adaptability and efficiency of AI agents in real-world applications.

The framework's ability to retain and build upon past experiences represents a significant step forward in making AI agents more adaptable and efficient. The researchers conclude that while the method is not a complete solution to continuous learning, it provides a valuable tool for improving agent performance through structured knowledge retention and refinement. The results suggest that WikiSkill could be a useful addition to AI development, particularly for applications requiring persistent learning and adaptation over time. The framework's success in improving performance across various benchmarks highlights its potential for broader implementation in AI systems. The researchers emphasize that while the method has limitations, it offers a practical and effective way to enhance AI agent capabilities through structured knowledge retention and iterative improvement. The results indicate that WikiSkill could be a valuable tool for developers looking to improve the adaptability and efficiency of AI agents in real-world applications. The framework's ability to retain and build upon past experiences represents a significant step forward in making AI agents more adaptable and efficient. The researchers conclude that while the method is not a complete solution to continuous learning, it provides a valuable tool for improving agent performance through structured knowledge retention and refinement. The results suggest that WikiSkill could be a useful addition to AI development, particularly for applications requiring persistent learning and adaptation over time. The framework's success in improving performance across various benchmarks highlights its potential for broader implementation in AI systems. The researchers emphasize that while the method has limitations, it offers a practical and effective way to enhance AI agent capabilities through structured knowledge retention and iterative improvement. The results indicate that WikiSkill could be a valuable tool for developers looking to improve the adapt,