Stereo matching networks based on deep learning can accurately estimate disparity, but the performance is not ideal in active stereo systems. In this paper, we propose HRASnet, an active stereo matching algorithm based on hierarchical recurrent neural networks, which generates accurate depth for a pair of active stereo images through an end-to-end neural network. We use the Siamese network connected by multi-resolution residual as the feature extractor to improve the matching effect of weak pixels. Local attention mechanism is used to enhance the location correlation of features, and multi-resolution correlation cost volume is constructed instead of 4D cost volume to calculate pixel similarity more efficiently. We introduce an iterative contrast reconstruction loss that overcomes the error dependence of noise and distance. We have conducted experiments on open passive stereo matching datasets and active stereo images, and the results show that our method has good matching effect.