As an innovative grid computing technique for sharing the distributed memory resources in a widearea network, RAM Grid exploits the free nodes to provide remote memory for others. The data intensive applications with massive random disk IO, such as the web server or database systems, have potential benefits from RAM Grid. We analyzed the trace collected from real servers, and proposed the key issues and solution of the caching RAM Grid system for such applications, to reduce the IO overhead. Furthermore, we proposed an autonomic cooperation mechanism to improve the performance through time correlation and content redundancy detection and elimination. We have verified the effectiveness of the proposed method through comprehensive trace-driven simulations.