In this study, a Ta 2 O 5 -doped HfO x (HfTaO x ) thin film was deposited by cosputtering to serve as the rectifying layer for HfO x -based resistive random-access memory (RRAM) with a final structure of Pt/HfO x /HfTaO x /TiN/SiO 2 /Si. Incorporating the appropriate proportion of lattice and nonlattice O in the rectifying layer enabled forming-free RRAM operation. Moreover, by modifying the compliance current and making use of the deep reset operation, multilevel resistance states were realized. In neuromorphic computing, when mimicking artificial synapses, potentiation and depression were successfully induced, and low nonlinearity was demonstrated, implying efficient weight modulation and reduced energy and time for neural network training. Software-comparable Modified National Institute of Standards and Technology (MNIST) handwritten digit database inference accuracy (97.54%) was achieved for an RRAM-based fully connected neural network with the HfTaO x rectifying layer.