Bayesian age-depth models are central to paleoclimate research, linking depth in natural archives to calendar age. When synchronizing variable data from different sites, inference in these models requires sampling from high-dimensional, often multimodal posteriors. Consequently, standard samplers become trapped in local optima, miss plausible chronologies, and bias downstream analyses. Here, we investigate two enhanced sampling approaches, On-the-fly Probability Enhanced Sampling (OPES) and Parallel Tempering (PT), and find that tempering only the likelihood components that drive multimodality can improve exploration. Our results suggest that OPES and PT combined with targeted tempering provide robust age-depth models with implications for a wide range of applications in Earth sciences, including paleoclimate and paleomagnetic reconstructions.