Multi-access Edge Computing (MEC) is a key technology in the fifth-generation (5G) network and beyond. MEC extends cloud computing to the network edge (e.g., base stations, MEC servers) to support emerging resource-intensive applications on mobile devices. As a crucial problem in MEC, service migration needs to decide where to migrate user services for maintaining high Quality-of-Service (QoS), when users roam between MEC servers with limited coverage and capacity. However, finding an optimal migration policy is intractable due to the highly dynamic MEC environment and user mobility. Many existing works make centralized migration decisions based on complete system-level information, which can be time-consuming and suffer from the scalability issue with the rapidly increasing number of mobile users. To address these challenges, we propose a new learning-driven method, namely Deep Recurrent Actor-Critic based service Migration (DRACM), which is user-centric and can make effective online migration decisions given incomplete system-level information. Specifically, the service migration problem is modeled as a Partially Observable Markov Decision Process (POMDP). To solve the POMDP, we design an encoder network that combines a Long Short-Term Memory (LSTM) and an embedding matrix for effective extraction of hidden information. We then propose a tailored off-policy actor-critic algorithm with a clipped surrogate objective for efficient training. Results from extensive experiments based on real-world mobility traces demonstrate that our method consistently outperforms both the heuristic and state-of-the-art learning-driven algorithms, and achieves near-optimal results on various MEC scenarios.