Memory networks are a general class of neural network models for NLU tasks introduced by Weston et al. in 2014 in the context of end-to-end trained QA systems. Given a question and some supporting facts or relevant information, the task is to generate or select an appropriate answer. The model stores these facts in a persistent memory and is trained to perform reasoning based on them to produce an appropriate response.
The first paper on this topic was titled Memory Networks by Jason Weston, Sumit Chopra, and Antoine Bordes, and can be found at http://arxiv.org/abs/1410.3916.
As QA tasks come in many varieties, memory networks offer a flexible and modular framework where facts stored in memories could range from text to images, and answers can be generated or retrieved from a set of candidates.