This paper presents a self-supervised dynamic learning approach for long-term high-fidelity image transmission through unstabilized multimode fibers (MMFs). The proposed method, called the multi-scale memory dynamic-learning network (MMDN), leverages a multi-expert framework to model both long- and short-term dynamics of the unstable MMF channel. The network parameters are dynamically updated over time in a self-supervised manner, by learning from currently predicted pseudo-labels to synthesize the optimal inverse transmission model for subsequent image inference. MMDN achieves adaptive and accurate tracing of the variations on the optical propagation model in MMF, enabling parallel transmission of 1024 spatial degrees of freedom through 1 km-length fibers with >99.9% accuracy for over 1000s duration. The high-fidelity performance enables efficient transmission of high-resolution video with several orders of throughput enhancement by using compressive encoding, showing the feasibility of long-term high-fidelity spatial transmission. The MMDN is designed to adaptively handle the time-varying optical propagation in long MMFs. The network is composed of multiple subnetworks with different memory scales, which are updated adaptively to capture the system variations. The MMDN is tested on 100m- and 1km-length MMFs, demonstrating high accuracy in the transmission of arbitrary optical fields. The results show that MMDN achieves significantly better performance than static neural networks, especially for long-distance transmission. The MMDN is also tested on high-level encoded random images, demonstrating its ability to recover non-binary patterns with high accuracy. The results show that MMDN can permit long-term spatial information transmission with high fidelity. The MMDN is also tested on long-term high-throughput video transmission through unstabilized MMFs. The results show that MMDN can achieve high-accuracy spatial transmission over long distances, enabling efficient video transmission with several orders of throughput enhancement. The MMDN is also tested on natural image transmission, demonstrating its ability to reconstruct natural images with high accuracy. The results show that MMDN can permit long-term spatial information transmission with high fidelity. The MMDN is also tested on scalability and generalization, demonstrating its ability to transmit specific types of natural images. The results show that MMDN can permit long-term spatial information transmission with high fidelity. The MMDN is also tested on the stability analysis of MMF channels, demonstrating its ability to handle the time-varying optical propagation in long MMFs. The results show that MMDN can permit long-term spatial information transmission with high fidelity.This paper presents a self-supervised dynamic learning approach for long-term high-fidelity image transmission through unstabilized multimode fibers (MMFs). The proposed method, called the multi-scale memory dynamic-learning network (MMDN), leverages a multi-expert framework to model both long- and short-term dynamics of the unstable MMF channel. The network parameters are dynamically updated over time in a self-supervised manner, by learning from currently predicted pseudo-labels to synthesize the optimal inverse transmission model for subsequent image inference. MMDN achieves adaptive and accurate tracing of the variations on the optical propagation model in MMF, enabling parallel transmission of 1024 spatial degrees of freedom through 1 km-length fibers with >99.9% accuracy for over 1000s duration. The high-fidelity performance enables efficient transmission of high-resolution video with several orders of throughput enhancement by using compressive encoding, showing the feasibility of long-term high-fidelity spatial transmission. The MMDN is designed to adaptively handle the time-varying optical propagation in long MMFs. The network is composed of multiple subnetworks with different memory scales, which are updated adaptively to capture the system variations. The MMDN is tested on 100m- and 1km-length MMFs, demonstrating high accuracy in the transmission of arbitrary optical fields. The results show that MMDN achieves significantly better performance than static neural networks, especially for long-distance transmission. The MMDN is also tested on high-level encoded random images, demonstrating its ability to recover non-binary patterns with high accuracy. The results show that MMDN can permit long-term spatial information transmission with high fidelity. The MMDN is also tested on long-term high-throughput video transmission through unstabilized MMFs. The results show that MMDN can achieve high-accuracy spatial transmission over long distances, enabling efficient video transmission with several orders of throughput enhancement. The MMDN is also tested on natural image transmission, demonstrating its ability to reconstruct natural images with high accuracy. The results show that MMDN can permit long-term spatial information transmission with high fidelity. The MMDN is also tested on scalability and generalization, demonstrating its ability to transmit specific types of natural images. The results show that MMDN can permit long-term spatial information transmission with high fidelity. The MMDN is also tested on the stability analysis of MMF channels, demonstrating its ability to handle the time-varying optical propagation in long MMFs. The results show that MMDN can permit long-term spatial information transmission with high fidelity.