Autoencoders have been widely used in video anomaly detection, and there are many variants. However, since these Autoencoders use frames as the input of the reconstruction network, they can only learn the pixel information in the frame and lack other key information. In addition, complete reconstruction of the frame will lead to memory consumption and waste of resources. In order to learn multi‐dimensional information and reduce memory usage, we propose to use the Object‐meta instead of video frames, and the Memory Search Guided Autoencoder with Memory Pools (MSGAE‐MP) to reconstruct. Every Object‐meta comes from the object and is composed of the type mask, position mask, optical flow, and pixels of the object. In this way, the multi‐dimensional information carried by the input can be strengthened. After generating Object‐meta, the MSGAE‐MP will use the high‐dimensional information to guide search in memory modules during the reconstruction, and it will construct multi‐level memory pools, so as to reconstruct Object‐meta in different dimensions. Experiments show that our method is feasible and has achieved excellent results.