In this work, a novel interface engineering method is proposed to address the relatively large cycle-to-cycle variability of the emerging metal-oxide resistive random access memory (RRAM) device technology. This is achieved by synthesizing the solution-processable graphitic nanosheet (reduced graphene oxide, rGO) with defects of a controllable amount and further integrating it into RRAM as an oxygen exchange layer (OEL). It is demonstrated that rGO-inserted RRAM exhibits reduced cycle-to-cycle variability in the SET switching as compared with one that has a conventional transition metal thin film as OEL. This is best attributed to the fact that our rGO thin film provides nearly the same amount of oxidation-prone atomic sites for each programming cycle. This study is expected to greatly advance the RRAM-based neuromorphic computing by paving a practically viable route to enhance the accuracy of the deep learning model.