Abstract: The increasing penetration of distributed new energy has caused the distribution network to face severe challenges such as over-limit voltage and increased power loss. For the Volt-Var control problem, deep reinforcement learning can effectively solve the shortcomings of traditional optimization methods in terms of model dependence and solution speed. However, the existing deep reinforcement learning methods have limited feature extraction capabilities and poor control effects when faced with complex scenarios of large-scale distribution networks. Therefore, this paper proposes a multi-agent deep reinforcement learning control strategy considering self-attention and temporal-memory. First, the volt-var control problem is modeled as a decentralized partially observable Markov decision process. Then, based on the self-attention encoder and temporal- memory neuron, four kinds of neural network structures are designed, including feature extraction network, auxiliary training network, improved policy network and improved value network. Next, self-supervised learning is introduced, and the centralized training with decentralized execution process of the proposed algorithm is described. Finally, a case study is carried out on the modified IEEE 141-bus distribution system. The experimental results show that the proposed control strategy can effectively extract state features, memorize temporal information, and identify key components, showing a superior control effect on voltage stabilization and loss reduction, while also boasting enhanced robustness, interpretability, and training stability.