The existing trust-based recommendation algorithms usually assume that users are homogeneous, and therefore can't fully mine the trust relationship information. Moreover, the lack of efficient model for integrating similar relationship and trust relationship greatly affects the accuracy and reliability of those models. To solve the issue, this paper first proposes a trust-based recommendation algorithm called Trust-PMF. It combines similarity with trust to build user' preference and selects the target user's neighbors. Then, the probability matrix factorization model is extended by integrating memory-based idea and trust information, and a dynamic adaptive weight is used to determine the degree of influence of each part to form a unified and efficient Trust-PMF model. Finally, experiment results on Filmtrust and Epinions data sets are presented to demonstrate that the proposed method outperforms the state-of-the-art methods.