Because of the fluidity and continuity of data stream, the knowledge embedded in stream data is most likely to be changed as time goes by. Thus, in most data stream applications, people are more interested in the information of the recent transactions than that of the old. This paper proposes a method for mining the frequent patterns in an arbitrary sliding window of data streams. As data stream flows, the contents of the data stream are captured with a compact prefix-tree by scanning the stream only once. And the obsolete and infrequent items are deleted by periodically pruning the tree. To differentiate the patterns of recently generated transactions from those of historic transactions, a time decaying model is also applied. Extensive simulations are conducted and the experimental results show that the proposed method is efficient and scalable, and also superior to other analogous algorithms.