MB-DPOP is an important complete algorithm for solving Distributed Constraint Optimization Problems (DCOPs) by exploiting a cycle-cut idea to implement memory-bounded inference. However, each cluster root in the algorithm is responsible for enumerating all the instantiations of its cycle-cut nodes, which would cause redundant inferences when its branches do not have the same cycle-cut nodes. Additionally, a large number of cycle-cut nodes and the iterative nature of MB-DPOP further exacerbate the pathology. As a result, MB-DPOP could suffer from huge coordination overheads and cannot scale up well. Therefore, we present RMB-DPOP which incorporates several novel mechanisms to reduce redundant inferences and improve the scalability of MB-DPOP. First, using the independence among the cycle-cut nodes in different branches, we distribute the enumeration of instantiations into different branches whereby the number of nonconcurrent instantiations reduces significantly and each branch can perform memory bounded inference asynchronously. Then, taking the topology into the consideration, we propose an iterative allocation mechanism to choose the cycle-cut nodes that cover a maximum of active nodes in a cluster and break ties according to their relative positions in a pseudo-tree. Finally, a caching mechanism is proposed to further reduce unnecessary inferences when the historical results are compatible with the current instantiations. We theoretically show that with the same number of cycle-cut nodes RMB-DPOP requires as many messages as MB-DPOP in the worst case and the experimental results show our superiorities over the state-of-the-art.