In this talk, we investigate the theoretical aspects of eliminating the correlation from network data when network nodes are memory-enabled. We prove that the well-known prefix-free universal compression algorithms, when used for correlation elimination of the relatively small network packets, suffer from an inevitable large redundancy. We propose network compression via memory as a compression-based solution that overcomes the aforementioned fundamental performance limit of universal compression algorithms for memory-enabled network nodes. We further characterize the performance improvements of the network compression solution and demonstrate its effectiveness when employed in wired and wireless network applications.