Many advanced machine learning applications require a limited computer to send requests to a server where the response is calculated and then returned. In some applications, this architecture is impossible either because of the urgent need for a response or the necessity for the agent to interact with its environment. My area of research is learning what are the important features in data input in order to complete a task and doing so with minimal memory and minimal latency. One application for this kind of research is artificially intelligent prosthetics. Imagine an amputee waiting for their arm to sync with a server before they were able to pick up a new object. That'd be crazy! Working with Dr. Patrick Pilarski at the University of Alberta, I am optimizing how a computationally-limited reinforcement learning agent copes with its limitation. Developing time and memory efficient methods by which a limited agent can represent its world and choose actions is the primary work in my thesis.