This paper presents a self-supervised dynamic learning approach for long-term, high-fidelity image transmission through unstabilized multimode fibers (MMFs). The method, termed Multi-Scale Memory Dynamic-Learning Network (MMDN), uses multiple networks with both long- and short-term memory to adaptively update and ensemble, achieving robust image recovery. The MMDN is trained using predicted pseudo-labels and dynamically updates its parameters over time, enabling accurate tracing of optical propagation model variations. Experiments demonstrate over 99.9% accuracy in transmitting 1024 spatial degrees of freedom over 1km length MMFs for over 1000 seconds. The approach also enables efficient transmission of high-resolution video with significant throughput enhancement, showcasing its potential for practical applications such as remote imaging and optical communication. The MMDN's ability to handle system drifting and environmental disturbances makes it a promising solution for long-term, high-fidelity spatial information transmission.This paper presents a self-supervised dynamic learning approach for long-term, high-fidelity image transmission through unstabilized multimode fibers (MMFs). The method, termed Multi-Scale Memory Dynamic-Learning Network (MMDN), uses multiple networks with both long- and short-term memory to adaptively update and ensemble, achieving robust image recovery. The MMDN is trained using predicted pseudo-labels and dynamically updates its parameters over time, enabling accurate tracing of optical propagation model variations. Experiments demonstrate over 99.9% accuracy in transmitting 1024 spatial degrees of freedom over 1km length MMFs for over 1000 seconds. The approach also enables efficient transmission of high-resolution video with significant throughput enhancement, showcasing its potential for practical applications such as remote imaging and optical communication. The MMDN's ability to handle system drifting and environmental disturbances makes it a promising solution for long-term, high-fidelity spatial information transmission.