This paper delves into the unexplored frequency-dependent characteristics of beyond diagonal reconfigurable intelligent surfaces (BD-RISs). A generalized practical frequency-dependent reflection model is proposed as a fundamental framework for configuring fully-connected and group-connected RISs in a multi-band multi-base station (BS) multiple-input multiple-output (MIMO) network. Leveraging this practical model, multi-objective optimization strategies are formulated to maximize the received power at multiple users connected to different BSs, each operating under a distinct carrier frequency. By relying on matrix theory and exploiting the symmetric structure of the reflection matrices inherent to BD-RISs, closed-form relaxed solutions for the challenging optimization problems are derived. The ideal solutions are then combined with codebook-based approaches to configure the practical capacitance values for the BD-RISs. Simulation results reveal the frequency-dependent behaviors of different RIS architectures and demonstrate the effectiveness of the proposed schemes. Notably, BD-RISs exhibit superior resilience to frequency deviations compared to conventional single-connected RISs. Moreover, the proposed optimization approaches prove effective in enabling the targeted operation of BD-RISs across one or more carrier frequencies. The results also shed light on the potential for harmful interference in the absence of proper synchronization between RISs and adjacent BSs.