Compressing real-time input through bandwidth constrained connections has been studied within robotics, wireless sensor networks, and image processing. When there are bandwidth constraints on real-time input the amount of information to be transferred will always be greater than the amount that can be transferred per unit of time. We propose a system that utilizes a local diffusion process and a reinforcement learning-based memory system to establish a real-time prediction of an entire input space based upon partial observation. The proposed system is optimized for dealing with multi-dimension input spaces, and maintains the ability to react to rare events. Results show the relation of loss to quality and suggest that at higher resolutions gains in quality are possible.