Expansion of the scale to the traditional collaborative filtering recommendation systems and changes of users’ interest bring problems of decreased accuracy and real-time responsiveness. Collaborative filtering recommender systems based on clustered users using K-Means Algorithm can solve these two problems in some extent, however, with a local optimum defects. Under the premise of ensuring the real-time responsiveness, AntClass algorithm applied to users is proposed to overcome the shortcomings of K-Means algorithm. This paper also proposed to take the users’ ratings as a data stream, and use the pyramid time frame for data preprocessing, thus it reflects the change of users’ interest with the time. As a result, AntClass algorithm and the data stream filtered by pyramid time frame were combined to form the AntStream algorithm in this article. The experiment result shows that AntStream algorithm has improved not only the real-time responsiveness and also the accuracy to a greater extent.