Reconfigurable intelligent surfaces (RISs) have become one of the key technologies in 6G wireless communications. By configuring the reflection beamforming codebooks, RIS focuses signals on target receivers. In this paper, we investigate the codebook configuration for 1-bit RIS-aided systems. We propose a novel learning-based method built upon the advanced methodology of implicit neural representations. The proposed model learns a continuous and differentiable coordinate-to-codebook representation from samplings. Our method only requires the information of the user's coordinate and avoids the assumption of channel models. Moreover, we propose an encoding-decoding strategy to reduce the dimension of codebooks, and thus improve the learning efficiency of the proposed method. Experimental results on simulation and measured data demonstrated the remarkable advantages of the proposed method.