MemNet is a deep persistent memory network designed for image restoration tasks such as denoising, super-resolution, and JPEG deblocking. It introduces a memory block with a recursive unit and a gate unit to explicitly mine persistent memory through adaptive learning. The recursive unit learns multi-level representations under different receptive fields, while the gate unit adaptively controls how much of the previous states should be reserved and how much of the current state should be stored. The memory blocks are densely connected to enhance information flow and improve performance. MemNet outperforms existing state-of-the-art models in all three tasks, demonstrating its effectiveness in handling long-term dependencies. The network is implemented with 80 convolutional layers, making it one of the deepest networks for image restoration. Comprehensive experiments show that MemNet achieves superior performance in image restoration tasks, with the same structure achieving state-of-the-art results across denoising, super-resolution, and JPEG deblocking. The model's ability to handle different levels of corruption with a single model highlights its robustness and effectiveness.MemNet is a deep persistent memory network designed for image restoration tasks such as denoising, super-resolution, and JPEG deblocking. It introduces a memory block with a recursive unit and a gate unit to explicitly mine persistent memory through adaptive learning. The recursive unit learns multi-level representations under different receptive fields, while the gate unit adaptively controls how much of the previous states should be reserved and how much of the current state should be stored. The memory blocks are densely connected to enhance information flow and improve performance. MemNet outperforms existing state-of-the-art models in all three tasks, demonstrating its effectiveness in handling long-term dependencies. The network is implemented with 80 convolutional layers, making it one of the deepest networks for image restoration. Comprehensive experiments show that MemNet achieves superior performance in image restoration tasks, with the same structure achieving state-of-the-art results across denoising, super-resolution, and JPEG deblocking. The model's ability to handle different levels of corruption with a single model highlights its robustness and effectiveness.