A microgrid is the small-scale power system consisting of Distributed Energy Resources (DERs), consumers and battery energy systems. It can generate, store and consume the power within a specific local region, and there are some advantages such as lower energy cost and carbon emission utilizing more renewable energy resources. Additionally, a multi-microgrid, which connects multiple microgrids, is expected to enhance stability and resiliency. By sharing electricity demand and capacity based on the situation of each microgrid, multi-microgrids can effectively support each other and the main grid in the event of contingencies. However, controlling power flow within a multi-microgrid is complex due to the uncertainty of demand and capacity conditions. Hence, a more robust and stable control method is required for the multi-microgrid. This study proposes demand and capacity sharing protocols for the multi-microgrid using deep learning, specifically Long Short-Term Memory (LSTM) networks. Firstly, demand and capacity sharing decisions and protocols are designed by integrating the LSTM-based prediction models and decision-making processes. Next, numerical experiments and analyses are conducted using real time-series data of demand and supply. Finally, the effectiveness of proposed demand and capacity sharing decisions and protocols is evaluated and discussed.