Cell-free massive multiple-input-multiple-output is promising to meet the stringent quality-of-experience (QoE) requirements of railway wireless communications by coordinating many successional access points (APs) to serve the onboard users coherently. A key challenge is how to deliver the desired contents timely due to the radical changing propagation environment caused by the growing train speed. In this paper, we propose to proactively cache the likely-requesting contents at the upcoming APs which perform the coherent transmission to reduce end-to-end delay. A long-term QoE-maximization problem is formulated and two cache placement algorithms are proposed. One is based on heuristic convex optimization (HCO) and the other exploits deep reinforcement learning (DRL) with soft actor-critic (SAC). Compared to the conventional benchmark, numerical results show the advantage of our proposed algorithms on QoE and hit probability. With the advanced DRL model, SAC outperforms HCO on QoE by predicting the user requests accurately.