In the expert recommendation algorithm of the Q & A community，the graph neural network mainly uses the interactive relationship between users and questions to build a model，and its model performance depends on the density of interactive data. So it is difficult to effectively represent and learn users and questions without inter? active information. This paper proposes an attention graph neural network expert recommendation method based on memory. Firstly，a multi-dimensional feature-oriented subnetwork is designed，and then a memory network is con? structed to store the similar questions answered by users for each question. At the same time，an attention mechanism is introduced between user representation and similar question representation，and the vector representation of new questions is constructed by fusing similar questions from different users′ perspectives，Finally，experts are recom? mended based on the representation of users and questions，which effectively improves the accuracy of expert recom? mendation. The proposed method is validated on the Q & A community data set and public data set，and its perfor? mance is improved compared with other similar models.