3D Gaussian Splatting (3DGS) has attracted attention as a novel view synthesis technique, but its high memory consumption limits its application on edge devices. Existing 3DGS compression methods focus solely on storage space, failing to address the critical bottleneck of rendering memory. In this paper, we present MEGS², a memory-efficient framework that jointly optimizes the total number of primitives and per-primitive parameters to achieve unprecedented memory compression. Specifically, we propose a unified soft pruning framework that uses lightweight, arbitrarily oriented spherical Gaussian lobes as color representations instead of memory-intensive spherical harmonic functions, and models both primitive and lobe pruning as a single constrained optimization problem. Experimental results show that MEGS² reduces static VRAM by 50% and rendering VRAM by 40% compared to existing methods, while maintaining comparable rendering quality.
