KnowAct-GUIClaw improves GUI agent accuracy and cross-platform adaptability

Researchers introduce KnowAct-GUIClaw, a framework addressing OpenClaw's limitations in cross-platform GUI interaction and self-evolution by unifying cognitive comprehension with operational execution. The system utilizes a pluggable GUI subagent equipped with an experience-attributable memory system and a self-evolving skill library to enhance task decomposition and tool call accuracy.

• The host agent leverages accumulated interaction experience for long-horizon task decomposition and allocation.
• A self-evolving skill library enables seamless cross-platform migration across Android, iOS, HarmonyOS, and Windows.
• GUIClaw with open-source Kimi-2.6 models achieves 64.1% on the MobileWorld benchmark, outperforming Seed-2.0-Pro and GPT-5.5.
• The framework's memory and skills are transferable across base models, improving performance by 8.5% with Kimi-2.6.

The authors consider this significant because it breaks through recursive self-improvement constraints, allowing the assistant to continuously store user profiles and feedback for better efficiency.