To deal with the time-varying and non-linear problems in near infrared (NIR) spectroscopy modeling, the recursive modeling algorithm has been introduced within a justin-time framework by a moving window. Recursive strategy is quite effective by adding of new samples and discarding oldest samples. For fermentation process, while moving window is adopted to facing the changing of target property, the initial database is expected to be a wide coverage to ensure the robustness of NIR model. To make a balance between robustness (more initial samples included) and adaptability (more impact of new samples), a modified recursive locally weighted modeling approach is proposed and applied in a Chinese yellow wine (CYW) fermentation process. Meanwhile, the distance measurement of the original recursive locally weighted algorithm is improves by taking target property into consideration. The proposed approach improves window moving strategy and distance measurement for NIR modeling, which can fully preserve the initial database and achieve high detection accuracy.