Stability is fundamental to filter design. It reflects the sensitivity of the estimate to initial conditions. In this paper, we focus on the stability of linear set-membership filters (SMFs), which has not yet been well understood. Considering linear time-invariant sampled-data systems, we explore the stability issue of the existing/classical SMFing framework and establish a new framework with guaranteed stability. First, we propose a new concept -the Observation-Information-Tower (OIT), which describes how the measurements affect the estimate in a setintersection manner. The OIT enables a rigorous stability analysis of the classical linear SMFing framework, where an explicit stability criterion is provided. It turns out that the classical framework requires the knowledge of the true initial range of the system state to guarantee stability, which is hence sensitive to initial conditions. To handle this problem, we establish an OITinspired filtering framework such that the stability is unaffected by the initial conditions. Under this stability-guaranteed framework, we develop a stable and fast constrained zonotopic SMF.