Sequential recommendation can capture dynamic interest patterns of users based on user interaction sequences. Recently, there has been interest in integrating the knowledge graph (KG) into sequential recommendation. Existing works suffer from two main challenges: a) representing each entity in the KG as a single vector can confound heterogeneous information about the entity; b) triple-based facts are modeled independently, lacking the exploration of high-order connectivity between entities. To solve the above challenges, we decouple the KG into two subgraphs, namely CRoss-user Behavior-based graph and Intrinsic Attribute-based graph (Crbia), depending on the type of relation between entities. We further propose a CrbiaNet based on the two subgraphs. First, CrbiaNet obtains behavior-level and attribute-level semantic features from these two subgraphs independently by different graph neural networks, respectively. Then, CrbiaNet applies a sequential model incorporating these semantic features to capture dynamic preference of the users. Extensive experiments on three real-world datasets show that our proposed CrbiaNet outperforms previous state-of-the-art knowledge-enhanced sequential recommendation models by a large margin consistently.