This paper proposes a compact 3D Gaussian Splatting (GS) SLAM system to address the high memory and storage costs of existing GS-based SLAM systems. The authors observe that the Gaussian ellipsoids generated by SLAM systems exhibit high geometric similarity, leading to redundancy and increased memory usage. To mitigate this, they introduce a sliding window-based masking strategy to remove redundant ellipsoids and a geometry codebook-based quantization method to compress Gaussian attributes. A local-to-global bundle adjustment method with reprojection loss is used for accurate pose estimation. The proposed system achieves faster training, rendering speed (226% increase), and lower memory usage (2.21× compression) while maintaining state-of-the-art scene representation quality. The system is evaluated on multiple datasets, demonstrating its effectiveness in dense visual SLAM tasks. The key contributions include a compact GS-based SLAM framework, a sliding window-based masking method, and a codebook-based quantization approach. The results show that the proposed method significantly improves efficiency and performance in real-time SLAM applications.This paper proposes a compact 3D Gaussian Splatting (GS) SLAM system to address the high memory and storage costs of existing GS-based SLAM systems. The authors observe that the Gaussian ellipsoids generated by SLAM systems exhibit high geometric similarity, leading to redundancy and increased memory usage. To mitigate this, they introduce a sliding window-based masking strategy to remove redundant ellipsoids and a geometry codebook-based quantization method to compress Gaussian attributes. A local-to-global bundle adjustment method with reprojection loss is used for accurate pose estimation. The proposed system achieves faster training, rendering speed (226% increase), and lower memory usage (2.21× compression) while maintaining state-of-the-art scene representation quality. The system is evaluated on multiple datasets, demonstrating its effectiveness in dense visual SLAM tasks. The key contributions include a compact GS-based SLAM framework, a sliding window-based masking method, and a codebook-based quantization approach. The results show that the proposed method significantly improves efficiency and performance in real-time SLAM applications.