Terahertz (THz) communications combined with reconfigurable intelligent surfaces (RIS) and massive multiple-input multiple-output (MIMO) technologies are considered key enablers for sixth-generation (6G) wireless networks. However, conventional hybrid beamforming optimization suffers from high computational complexity due to the large number of antennas and RIS reflecting elements. In this paper, a deep reinforcement learning (DRL)-based low-complexity hybrid beamforming framework is proposed for RISassisted THz massive MIMO systems. The proposed approach jointly optimizes the analog beamformer, digital beamformer, and RIS phase shifts while significantly reducing computational complexity. The DRL agent learns the optimal beamforming strategy according to channel state information (CSI), signalto-noise ratio (SNR), and RIS configurations. Simulation results demonstrate that the proposed scheme achieves higher spectral efficiency, improved energy efficiency, and lower computational complexity compared with conventional optimization methods.