Abstract: High-precision full reactor simulation is extremely important for reactor design and safety analysis. In recent years, three-dimensional (3D) full reactor transport calculation becomes one of the key development directions of deterministic calculation in reactor physics. 3D full reactor transport calculation requires the calculation program to have strong anisotropy, fine geometric processing and large-scale parallel computing capabilities. The method of characteristics (MOC) is one of the excellent implementations of the next generation of neutronic deterministic calculations due to its powerful geometric processing capabilities and natural parallelism. However, the huge requirements of the MOC for computing time and space storage are also important factors restricting development. The main technical routes of MOC include 2D/1D method and direct 3D method. Among them, the direct 3D MOC has powerful geometric processing capabilities, but its performance requirements for computers are harsh, especially the large storage scale will limit the applicability of the machine. To this end, a characteristic tracking method based on hierarchical modeling tree and improved R-function method was proposed, and on this basis, on-the-fly generation of characteristic information technology was adopted, which could effectively reduce computation time and spatial storage. The numerical results of the Takeda and C5G7 benchmarks show that the method have good computational accuracy compared to the reference solution in the calculation of eigenvalues and other benchmarks. The proposed efficient characteristic tracking and on-the-fly generation technology can significantly reduce spatial storage. After increasing the time cost by about 45% and 23% respectively, the storage of the two benchmarks can be reduced by up to 99.51% and 99.92%, respectively. It is proved that this method can greatly improve the machine applicability of the direct 3D MOC.