Predictive coding posits that the brain is constantly forming expectations of what will happen next in a sensory stream and adjusting neural activity towards these expectations. We build on top of the classic Rao & Ballard implementation of predictive coding and more recent work showing how supervised learning can be performed using predictive coding by targeting a 1-hot vector of neural activity at the top of the network. Here, instead of assuming access to labels, which is biologically unrealistic, we aim to explore a variety of potential neural target patterns drawn from stored memories. This approach is motivated by questions of how hippocampus and cortex interact while performing tasks, recent neuroscientific theories/evidence showing that neurons may adjust their neural activity to align with target activity, and renewed interest in memory mechanisms, in the machine learning community. This work directly addresses biological learning rules by virtue of the Hebbian learning present in predictive coding systems, and continual learning due to the evolving memory cache. This project is a new collaboration between the labs of Professor Richard Zemel at Columbia University and Blake Richards at McGill University, and will advance ARNI’s vision by translating recent advances in neuroscience into scalable machine learning systems that can serve as simulations of the brain.