In the operation of logistics cargo, temporal efficiency in handling logistics is a critical issue. One solution to enhance this efficiency is the deployment of multiple autonomous ground vehicles (AGVs). This paper proposes a reinforcement learning approach based on Long Short-Term Memory (LSTM) for multi-agent path planning in the logistics cargo environment. When AGVs are treated as moving obstacles to each other, the application of LSTM allows for path planning that aptly addresses changes over time within the environment. Additionally, to solve the well-known problem of the sparse reward in pathfinding, we propose a reinforcement learning architecture for multi-agent path planning that uses the path planning results of a single agent, guided by Q-learning, as the guide path. Furthermore, we have established state variables independent of the number of agents by setting interest window-based state variables, and introduced revisit rewards, effectively resolving the issue of local minima caused by repetitive movement-avoidance behavior between agents. Simulation results, emulating a real-world logistics warehouse environment, demonstrate the proposed technique’s capability for effective multi-agent path planning in such settings.