A generic memory network's architecture can be decomposed into four parts: a Question Module, an Input Module, a Memory Module, and an Output Module. As is common practice in neural networks, information passes from one module to the other through dense vectors/embeddings, making the parameters of the model end-to-end trainable using gradient descent:

The working of this model is as follows:

• The Input Module receives multiple facts and encodes each of them in vectors.
• The Question Module, similar to the Input Module, is responsible for encoding the question in a vector.
• The Memory Module receives the encoded facts (from the Input Module) and the encoded question (from the Question Module), and performs a soft attention mechanism on the facts to figure out their relevance to the question. The result of the attention is a context vector for the given question, encoding the question, as well as all the contextual information required to answer it.
• The Output Module receives the context vector and is responsible for producing an answer in the desired format. This could mean the selection of an appropriate response from a candidate set, the prediction of answer spans, or the generation of a token-by-token response.