Virtual walkthrough technology has a wide range of applications in the metaverse, and it is also a basic function of most 3D video games. Specifically, it means that the user controls the virtual perspective, roams in the 3D virtual space, and is used to visually feel the transfer of the body position in the virtual space. We propose a novel scheme: in-situ pose classification via long-term memory augmented networks. After testing, our new method achieves an accuracy of 98.12% with a latency of 181ms, outperforming other recent in-situ pose classification techniques.