To address the problems of traditional buffer overflow vulnerability repair methods, such as reliance on fixed templates and difficulty in handling complex variants and runtime discrepancies, this study proposes an automated repair framework based on the micro-patch adaptive optimization algorithm (AOMP). By conducting structural modeling on the contextual features of vulnerabilities and applying the Q-learning algorithm to optimize strategies within a predefined patch-primitive library, the framework automatically generates optimal micro-patches for different vulnerability states. Experiments conducted on multi-layer datasets show that AOMP achieves a repair success rate of 98.2% on basic overflow samples and 93.3% on complex variant vulnerabilities, with 100% defense effectiveness achieved in real-world vulnerability scenarios. Meanwhile, the average performance overhead is only 4.8%, which is 55% lower than that of the traditional tool TuxFix. The results verify that the proposed framework can dynamically adjust patch structures according to vulnerability characteristics, enabling adaptive repair of unknown variants and providing technical support for system security.