Cross-laminated timber (CLT) joints are essential in ensuring co-working performance between CLT structural members and dissipating seismic energy when structures suffer earthquakes. The variation of mechanical properties of CLT joints would further affect the whole structural safety. Thus, evaluating and predicting the hysteretic behaviors of CLT joints under cyclic loading is significant. However, complex structures combined with multiple materials’ parameters increase the difficulty and effort of existing destructive mechanical experiment methods and finite element methods (FEM). The authors proposed a non-destructive data-driven method based on a recurrent neural network (RNN) to provide accurate and efficient hysteretic loading behavior prediction for CLT joints in this research herein. Multi-dimensional eigenvector is constructed by fusing damage-sensitive features extracted from lead zirconate titanate (PZT)-enabled sensing signals and material parameters and then training a gated recurrent unit (GRU) model. The prediction performance was assessed by statistical indexes (MAE, RMSE, and R2) and mechanical property indexes, respectively. Prediction results demonstrate that features extracted from PZT-enabled sensing signals could improve the predicting accuracy with R2 increasing from 0.94 to above 0.97, errors of mechanical property indexes decreasing from 0.20 to 0.14. The proposed method could predict the hysteretic behavior of the CLT joint with high accuracy and consistency with the test results in mechanical properties, which perform better than other models (LSTM, BiLSTM, and BiGRU). Overall, the significant contribution of the proposed method is that it can be applied to predict the ongoing hysteretic curves using available non-destructive sensing data and recurrent neural networks, which can provide an alternative of the destructive and time-consuming way to learn the development of hysteretic behavior of CLT joints under cyclic loading.