Relay mechanisms are an important part of communication systems and, therefore, naturally occurring molecular communication (MC) links. Multiple techniques have been proposed for designing MC relay-aided setups, assuming synchronous operation and perfect timing during the decoding process. In this paper, we propose using a reinforcement learning (RL)-based synchronizer to continually adapt a decoding threshold and detect transmitted synchronization frames in a dynamic MC environment. We implement our approach in a two-hop MC link model with mobility and show its advantages compared to filter-based maximum likelihood (ML) synchronization. Thereby, we utilized a macroscale, air-based MC testbed for the experimental determination of the channel impulse response (CIR) for a more realistic channel model. Our simulation results exhibit the potential of an RL-based synchronizer with a similarly high detection rate, a false positive rate one order of magnitude lower, and a misalignment several bit times lower compared to the state of the art.