Vision-and-Language Navigation (VLN) refers to the process by which robots navigate visual environments by following natural language instructions from humans. A key challenge in VLN, particularly for long-distance navigation, is efficiently maintaining the robot’s memory of historical navigation steps. To address this, we propose the Memory-Driven Querying Transformer (MDQT), a new model that leverages a learnable fixed-length memory vector to encode historical information. At each step, the memory vector interacts with both linguistic and visual information via a Querying Transformer (Q-Former). MDQT then predicts the next action and updates the memory vector using the Memory Update Module (MUM). The model is pre-trained on three tasks and fine-tuned within simulation environments. Experimental results demonstrate that MDQT achieves a 53% Success Rate (SR) on the Room-to-Room (R2R) dataset, outperforming baseline models. Ablation studies further validate the effectiveness of our pre-training tasks and MUM in enhancing navigation performance. Additionally, we transfer the model from simulation environments to a physical robot, enabling it to navigate in real-world environments.