We develop a geometric reassembly algorithm that composes a 3D model from its fragments. One important application of this work is skull completion and modeling in archeology and forensics. Our reassembly algorithm employs a scale-space representation of shape based on the heat kernel, which only depends on the intrinsic geometry of the surfaces. Partial matching can then be conducted effectively. The entire assembly pipeline has three steps: (1) fragment-template matching based on heat-kernel; (2) matching refinement based on RANSAC and assembly computation; (3) assembly refinement using least square transformation error (LSTE) of break-curves. The main contribution of this paper is presenting novel algorithms for the first two steps. Experimental results on scanned skull fragments demonstrate the efficacy and robustness of our algorithm.