Lifelong learning is a learning paradigm in which learning agents receive new tasks sequentially, in a lifelong manner. When a new task arrives, the knowledge from all previous tasks is leveraged in order to positively bias the learning of the new task. This is an attractive ability for intelligent agents to have as it mitigates the problem of having to learn new tasks from scratch each time, as well as alleviating the need for large labelled datasets that are both scarce and of high cost. Another attractive ability is that of zero-shot learning, where a task is solved without ever being explicitly trained to do so. Zero-shot learning is complemented by learning since lifelong learning is concerned with accumulating prior knowledge over time -- resulting in strong priors and inductive biases that can aid in the goal of zero-shot learning. We propose to explore a novel framework for lifelong zero-shot learning with the aim of learning how to utilize arbitrary prior knowledge.