Predominant resources for execution of any application are computational power and memory. On one side, computational power has grown many folds faster than memory capacity. On the other side, application's memory requirements have kept on increasing from time to time. Application?s minimum memory requirement influences job scheduling decision in grid. But once the application starts executing it faces memory pressure i.e. increase in memory requirement. This could be handled by remote memory paging - moving pages from memory loaded machine to remote machine with unused memory. Highly unpredictable network latency in grid has direct impact on the remote memory access latency. The idea of prediction and prefetching can be adapted to reduce this latency. Profile based and Markov based prediction models are explored in this paper. The experiments on memory intensive applications show that the Markov based model has better accuracy and profile based prediction provide good coverage.