For reasoning and reading comprehension tasks, the Dynamic Memory Network, which included attention mechanism, shows strong reasoning and understanding ability for some kinds of problems. But there is more to improve. We believe that in the original iteration structure of the DMN model, even after many iterations, some important information will be missed, or useless information will be considered too much. Based on the above points, we refer to DMN plus model on the basis of DMN model and make improvements in the extraction of attention features with a Two-Pass Attention Mechanism and made some change in output module so that it can improve the accuracy of some kinds of tasks.