This work presents an LSTM-based framework for predicting the temporal evolution of threshold-voltage (Vt) distributions and retention-induced lifespan in 3D NAND flash memory. Trained on large-scale simulation data, the model learns percentile-wise Vt trajectories and reconstructs the full distribution to estimate page-level lifetime under arbitrary failure criteria. The results demonstrate accurate prediction of key distribution metrics and reliable lifespan estimation suitable for practical reliability assessment.