The paper introduces KnowAct-GUIClaw, a Know-Route-Act-Reflect framework intended to improve OpenClaw’s cross-platform GUI interaction and continual self-improvement. It combines long-horizon task decomposition, an experience-attributable memory system, a pluggable GUI subagent, and a self-evolving skill library. The system targets Android, iOS, HarmonyOS, and Windows. According to the abstract, a configuration using open-source Kimi-2.6 reached 64.1% on the long-horizon MobileWorld benchmark, reportedly outperforming listed agent frameworks and models including Seed-2.0-Pro and GPT-5.5. Transferring memory and execution skills across base models produced an 8.5% improvement with Kimi-2.6.
No heat snapshots are available in the last 24 hours.