Servlet cache can effectively improve the throughput and reduce response time experienced by customers in servlet container. An essential issue of servlet cache is cache replacement. Traditional solutions such as LRU, LFU and GDSF only concern some intrinsic factors of cache objects regardless of associations among cached objects. For higher performance, some approaches are proposed to utilize these associations to predict customer visit behaviors, but they are still restricted by first-order Markov model and lead to inaccurate predication. In this paper, we describe associations among servlets as sequential patterns and compose them into pattern graphs, which eliminates the limitation of Markov model and achieve more accurate predictions. At last, we propose a discovery algorithm to generate pattern graphs and two predictive probability functions for cache replacement based on pattern graphs. Our evaluation shows that this approach can get higher cache hit ratio and effectively improve the performance of servlet container.