An adaptive feature-matching network is proposed to solve the common problem of object occlusion in object tracking. By calculating the pixel-level similarity between the query and memory frames， the network encodes the similarity relationship between an object and its background and obtains a pixel-level similarity matrix. By separating the query and memory frames， the network calculates the multi-dimensional similarity to focus on more areas in the query frame and adaptively weighs the memory frame through the calculated similarity matrix to improve the accuracy and robustness of object tracking. Additionally， the feature memory network selects and saves the memory frames， provides additional apparent information for feature matching， and allows the network to implicitly learn the moving trend of an object to achieve better tracking results. Experimental results show that this method performs well on GOT-10k， LaSOT， and other datasets. On GOT-10k datasets， compared with the STMTrack algorithm， the value of the proposed algorithm is improved by 1.8%. The visualization results show that the proposed algorithm is more robust in meeting the challenges of object occlusion and disappearance.