Abstract: To meet the rapid evolution of artificial intelligence (AI) algorithms, the demand for computing resources is growing exponentially, posing significant challenges for deploying AI models on hardware. Memristor-based computing in memory accelerators offer a promising solution to the energy efficiency and latency issues in large-scale AI model deployment—by performing computation directly within the memory cells where data are stored. This approach eliminates the frequent data transfers between processing units and memory units inherent in the von Neumann architecture, thereby greatly reducing both time and energy consumption. In recent years, research in this field has progressed rapidly, and memristor technology is undergoing a critical transition from proof-of-concept to commercialization, with prototype systems already capable of accelerating AI model inference across various application scenarios. This article systematically reviews memristor devices and crossbar arrays, system architectures, software tools, representative applications, and development trends, and highlights the key technical challenges that remain to be addressed.