Intraoperative patient motion degrades static registration. We propose an adaptive AR guidance system with feedback-driven registration updates. The system employs particle filter-based motion compensation and multi-scale ICP refinement. In 10 cadaver experiments with head motion up to ±20 mm, average TRE remained under 1.5 mm, compared with 2.7 mm in static ICP. Accuracy improved by 43%, and frame rate stayed at 26 fps. The system enables real-time adaptation in dynamic surgeries.