In this article, we propose a multi-agent deep reinforcement learning (MADRL) framework to train a multiple access protocol for downlink low earth orbit (LEO) satellite networks. By improving the existing learned protocol, emergent random access channel (eRACH), our proposed method, coined centralized and compressed emergent signaling for eRACH (Ce2RACH), can mitigate inter-satellite interference by exchanging additional signaling messages jointly learned through the MADRL training process. Simulations demonstrate that Ce2RACH achieves up to 36.65% higher network throughput compared to eRACH, while the cost of signaling messages increase linearly with the number of users.