VisionHOPE is the official PyTorch implementation of "VisionHOPE: Visual Backbones as Self-Modifying Learning Systems" (arXiv 2609.33325), a visual backbone built on self-referential nested learning (SRNL) from the Nested Learning paper. The idea is that what the model remembers and how it learns evolve together while it processes an image. The VisionHOPE operator uses five coupled memories that store content, generate key and value representations, and govern learning rate and retention. A soft cap on injection and a spectral clamp are meant to keep the memory recurrences non-expansive, and four directional scans use row- and column-aligned chunks, followed by spatial restoration and learned channel-wise fusion. The hierarchical Tiny, Small and Base models are reported at 26.6M, 52.9M and 91.2M parameters, with ImageNet-1K top-1 accuracy of 84.1, 85.2 and 85.6 percent at 224x224. With Mask R-CNN on COCO val2017, the 1x schedule gives box AP of 47.9, 49.5 and 50.5 for T, S and B, and the 3x schedule gives 49.4 and 50.5 for T and S. With UPerNet on ADE20K at single scale, mIoU is 49.4, 50.3 and 51.8. The repository includes training and evaluation scripts for all three tasks, standalone SRNL, VisionHOPEOperator and VisionHOPEBlock modules, a fast-inference mode with a prepare_for_inference helper, and complexity and efficiency measurement commands. It requires Linux, an NVIDIA GPU, CUDA 12.1 and PyTorch 2.1, and its CUDA extensions compile on first use. Pretrained checkpoints are on Hugging Face, GitHub Releases and Baidu Netdisk. The code is MIT licensed.