Abstract: To address the challenge that the transverse and longitudinal failures of composite materials are dominated by distinct components, and inconsistencies in the convergence level and rate of individual mechanical property parameter predic-tion tasks during regression modeling result in constrained accuracy of synchronous modeling, this study presents the Dynamic Memory-Weight Algorithm (DMWA) for guiding the multi-parameter performance prediction of compo-site materials. Firstly, a long-term memory mechanism is proposed, which utilizes training history information to gener-ate long-term memory weights that reflect the convergence level of each task. Secondly, short-term memory weights are generated using real-time state information of individual parameter prediction tasks to characterize the convergence rate; the long-term and short-term memory loss weights are fused to balance the convergence level and rate of different tasks, thereby mitigating the degradation of prediction accuracy induced by conflicts among various performance parameter prediction tasks during training. Finally, case validation and analysis are performed by predicting the mechanical prop-erties (X_T、E_11T、Y_T、E_22T、G₁₂、S₁₂) using the experimental dataset of X850 carbon fiber/epoxy resin composite unidirectional laminates. The analysis results demonstrate that regarding prediction accuracy, the mean prediction error of each mechanical property parameter for the model trained with DMWA is below 5%, and the coefficient of determination (R²) values all exceed 0.90, exhibiting a significant improvement relative to the PSO-BPNN and the UW-BPNN. With respect to computa-tional efficiency, compared with single-task models that independently model each prediction task, DMWA achieves an 85% improvement in computational efficiency, effectively reducing the computational and temporal costs required for acquiring multiple performance parameters. DMWA efficiently resolves the task conflict problem in multi-task perfor-mance prediction of composite materials, strikes a balance between prediction accuracy and computational efficiency, and offers a viable approach for the rapid acquisition of composite mechanical properties. The Dynamic Memory-Weight Algorithm provides a new approach for the high-precision and high-efficiency acquisition of multiple composite material property parameters.