This document summarizes research on developing a unified framework for learning representations from EEG data during working memory tasks. Fifteen participants completed a Sternberg working memory task while EEG was recorded. The EEG data was preprocessed, transformed into images representing spectral properties over time, and analyzed using convolutional neural networks and recurrent neural networks. Classification results showed that incorporating temporal information, such as with LSTMs or 1D convolutions, improved accuracy over max pooling across time. Learned representations revealed frequency selectivity and links to known electrophysiological markers of cognitive load. Future work will explore applications to brain-computer interfaces.