The open hole well area located in the eastern part of the Yanchang Oilfield has an earlier logging time. There are only three logging data in this area: SP, GR and R2.5. Due to the lack of logging curves such as AC and RT, refined research on reservoir geology is difficult to meet. The region has the characteristics of long development time and low single well production, so it is infeasible and uneconomical to carry out logging again. Long short-term memory (LSTM) model can be used to reconstruct missing logging curves. It is an economical and effective method suitable for stratigraphic logging sequence data. However, the application effect of this model in the eastern part of Yanchang Oilfield is poor. The reason is that the loess layer overlying the shallow oil reservoirs in the eastern region has significant interference with logging data signals. To address this issue, the LSTM model was constrained by geological stratification and a new model was established to achieve curve reconstruction. The logging data of each well with a length of 6 layers is intercepted by layered data as sample data. This method not only retains the advantages of LSTM model for sequential data processing, but also avoids the interference of overlying loess layer logging data on the model. Open hole wells with complete logging data were utilized for model training and validation. The accuracy of LSTM logging curve reconstruction considering geological stratification constraints was discussed. The results indicate that by introducing geological stratification constraints, the accuracy of model reconstruction logging curves is higher. The optimized model was used to reconstruct AC and RT curves for 50 wells in the open hole area with only GR, SP, and R2.5 curve data. Secondary interpretation work was carried out on 142 perforated sections of 50 wells, and the coincidence rate of interpretation results compared to oil testing conclusions reached 89.4%. The practicality and effectiveness of this method in reconstructing logging curves have been verified.