Reconfigurable Intelligent Surfaces (RIS) have emerged as a promising technology to enhance wireless communication systems by programmatically steering incident signals. However, the joint optimization of the precoding matrix at the Base Station (BS) and the phase-shifting matrix at the RIS elements remains a significant challenge due to the highly non-convex nature of the optimization problem and the diversity of communication environments. Traditional optimization methods suffer from high complexity, while deep learning (DL) approaches often lack robustness in various scenarios. To address these issues, the paper introduces a gradient-based manifold meta learning method (GMML), which does not require pre-training and offers strong robustness for RIS-aided communications. GMML combines meta learning and manifold learning to improve spectral efficiency and reduce the overhead of high-dimensional signal processing. Unlike traditional DL methods that directly use channel state information (CSI) as input, GMML feeds gradients of the precoding and phase-shifting matrices into neural networks. A differential regulator is designed to constrain the phase-shifting matrix of the RIS. Numerical results show that GMML improves spectral efficiency by up to 7.31% and speeds up convergence by 23 times compared to traditional approaches. It also demonstrates superior robustness and adaptability in dynamic settings, making it suitable for real-world applications. The proposed method is validated through simulations, showing its effectiveness in both perfect and imperfect CSI setups, and its ability to handle dynamic scenarios with high mobility.Reconfigurable Intelligent Surfaces (RIS) have emerged as a promising technology to enhance wireless communication systems by programmatically steering incident signals. However, the joint optimization of the precoding matrix at the Base Station (BS) and the phase-shifting matrix at the RIS elements remains a significant challenge due to the highly non-convex nature of the optimization problem and the diversity of communication environments. Traditional optimization methods suffer from high complexity, while deep learning (DL) approaches often lack robustness in various scenarios. To address these issues, the paper introduces a gradient-based manifold meta learning method (GMML), which does not require pre-training and offers strong robustness for RIS-aided communications. GMML combines meta learning and manifold learning to improve spectral efficiency and reduce the overhead of high-dimensional signal processing. Unlike traditional DL methods that directly use channel state information (CSI) as input, GMML feeds gradients of the precoding and phase-shifting matrices into neural networks. A differential regulator is designed to constrain the phase-shifting matrix of the RIS. Numerical results show that GMML improves spectral efficiency by up to 7.31% and speeds up convergence by 23 times compared to traditional approaches. It also demonstrates superior robustness and adaptability in dynamic settings, making it suitable for real-world applications. The proposed method is validated through simulations, showing its effectiveness in both perfect and imperfect CSI setups, and its ability to handle dynamic scenarios with high mobility.