Abstract: In oil and gas drilling engineering, dynamic optimization of drilling parameters is a core technical step for improving drilling efficiency and reducing operational costs. However, existing full-well-section parameter optimization methods generally adopt a fixed operation mode with continuous tracking throughout the entire process, lacking flexible start-stop adaptability, making it difficult to meet the engineering needs of “on-demand optimization and real-time solving” in field operations. In response, a memoryless drilling parameter dynamic optimization method for the full well section was proposed based on a long short-term memory network (LSTM). This method overcame the difficult convergence of traditional recurrent neural networks (RNNs) in long-sequence training. Based on LSTM, a high-precision mapping model was constructed between the time-series data stream of drilling parameters and the rate of penetration (ROP), and parameters were optimized with maximizing ROP as the objective. By constructing a multi-dimensional time-series dataset of drilling parameters, the method enabled real-time parameter optimization for any well section without inheriting the state information of previous well sections. It featured on-demand start-stop and independent solving, and did not rely on historical optimization results. Experimental results showed that under complex formation conditions, the proposed method improved the parameter prediction accuracy by approximately 90% compared with the traditional RNN model. In comparison with the traditional mode of continuous tracking throughout the entire process, it offers greater engineering convenience and field practicality, and can provide reliable technical support for the engineering implementation of intelligent drilling systems.