The five-character quatrain is the treasure of Chinese traditional literature, rendering people a unique language aesthetic and aesthetic experience. The process of machine-generated quatrain has a positive exploration of machine acquisition on human language. Inspired by the rhythm and antithesis features of poetry language itself, we trained language memory model on poetry datasets and couplet datasets. The model consists of a semantic model and a textual rule model. The semantic model uses a one-dimensional convolution network to extract the semantic features of poetry and learn the semantic information of the poetry. The word model uses an encoder-decoder model with attention mechanism to learn the isocolon features of poetry writing. The experimental results show the language memory model can generate poems that conform to the rules of poetry and our aesthetics.