Soft sensors have been widely used in chemical processes to predict values of difficult-to-measure process variables online. If the relationship between explanatory variables X and an objective variable y is changed by catalyst deterioration, change of product and so on, prediction accuracy of a soft sensor is reduced. This problem is called degradation of a soft sensor model. To overcome the degradation, many adaptive soft sensors have been proposed. In this paper, we aim to improve prediction accuracy of just-in-time (JIT) models. JIT models are constructed with only data close to a query or with all data having weights according to similarity with a query. If the type of degradation is shift of y-value, prediction accuracy of JIT models is reduced since data with similar X-values but different y-values are mixed in database and the relationship between X and y is not consistent. To resolve this problem, we propose to update database based on not only X-distance but also y-distance. The updated database is called JIT database. When a y-value is measured and a datum of X and y is obtained, data whose X-distance is low and y-distance is high from the datum are moved from JIT database to original database, and data whose X-distance and y-distance are both low are moved from original database to JIT database. To evaluate the performance of the proposed method, we used fifteen types of simulation data containing five types of state transition (Y-shift, X-shift, Slope-change, Y-shift + Slope-change and X-shift + Slope-change), and three types of transition speed (Instant, Rapid and Gradual). By using the proposed method, improvement of prediction accuracy of JIT models was achieved for all types of simulation data.