Abstract: Memory mechanism research is a core topic in computational neuroscience.While existing studies have revealed the importance of synaptic plasticity,signal transmission,and other factors in memory formation within neural networks,accurately replicating these complex biological mechanisms in computational models remains a significant challenge.Traditional graph-theoretical models can describe the topological structure of neural networks,but their static nature and centralized information processing cannot fully capture the dynamic,distributed,and decentralized characteristics of biological neural networks.Therefore,there is an urgent need for a memory model based on dynamic directed graphs that better aligns with the features of biological neural systems.The model is based on memory trace theory and employs a sparsely connected directed graph network structure,allowing autonomous decision-making through local information to achieve decentralized parallel processing.The model innovatively introduces a variable resistance structure,dynamically adjusting resistance values to simulate synaptic plasticity in neurons.Additionally,a resource-competition-based path reinforcement mechanism is used to mimic the memory process in biological neural networks.Experimental results show that the model consistently achieves memory functionality across various network scales and topologies,and the memory capacity increases approximately linearly with network width,exhibiting characteristics like those of biological neural networks.Compared to existing mainstream models,the proposed model demonstrates significant advantages in resource efficiency,memory capacity,and network scalability,providing a solid theoretical foundation for the design of brain-like computing systems and the development of intelligent systems.

Key words: Memory modeling, Distributed algorithm, Directed graph