The recent advances in hardware and software has led to development of applications generating a large amount of data in real-time. To keep abreast with latest trends, learning algorithms need to incorporate novel data continuously. One of the efficient ways is revising the existing knowledge so as to save time and memory. In this paper, we proposed an incremental algorithm for Bayesian network structure learning. It could deal with high dimensional domains, where whole dataset is not completely available, but grows continuously. Our algorithm learns local models by limiting search space and performs a constrained greedy hill-climbing search to obtain a global model. We evaluated our method on different datasets having several hundreds of variables, in terms of performance and accuracy. The empirical evaluation shows that our method is significantly better than existing state of the art methods and justifies its effectiveness for incremental use.