The memory cost and data efficiency of Reinforcement Learning (RL) models are critical factors in obtaining optimal performance. While some of these problems are addressed by existing techniques like Generative Replay and Prioritized Experience Replay, they frequently cause instability and excessive memory usage. This work presents a elevated architecture that uses input denoising to improve data efficiency and ensures high-quality data usage without big buffers. Our design uses four Model, respectively Synthesis Model, Denoise Model, Action Model, and Variance Model models into the actor framework. To create well-rounded actions, the Denoise Model cleans the input data, the Synthesis Model returns the ideal actions, the Action Model takes historical context into account, and the Variance Model synthesizes outputs.