<p>Meshless methods have achieved considerable success in solid mechanics by effectively handling large deformations and enabling adaptive refinement without the need for mesh generation. However, their convergence and accuracy can be sensitive to patch configurations and local approximations, typically requiring problem-specific tuning. While recent advanced neural network-based solvers offer a promising alternative, they may still struggle to accurately capture local features in complex domains. This study introduces a neural network-augmented meshless method for solid mechanics. Specifically, light-weight neural networks are employed for adaptive patch shape description and local weight functions within an energy-based training scheme. Furthermore, nodal enhanced neural network bases are introduced as local approximations to capture local mechanical behaviors accurately. To validate the proposed method, various three-dimensional numerical examples in linear elasticity, hyperelasticity, and elastoplasticity are presented. The comparative results indicate that the proposed framework achieves higher accuracy and faster convergence with limited network parameters and solution iterations.</p>