The increasing complexity of modern technical systems can exacerbate model uncertainty in model-based control, posing a great challenge to safe and effective system operation under closed loop. Online learning of model uncertainty can enhance control performance by reducing plant–model mismatch. This article presents a learning-based stochastic model predictive control (LB-SMPC) strategy for reference tracking of stochastic linear systems with additive state-dependent uncertainty. The LB-SMPC strategy adapts the state-dependent uncertainty model online to reduce plant–model mismatch for control performance optimization. Standard reachability and statistical tools are leveraged along with the state-dependent uncertainty model to develop a chance constraint-tightening approach, which ensures state constraint satisfaction in probability. The stability and recursive feasibility of the LB-SMPC strategy are established for tracking time-varying targets, without the need to redesign the controller every time the target is changed. The performance of the LB-SMPC strategy is experimentally demonstrated on an atmospheric pressure plasma jet (APPJ) testbed with prototypical applications in plasma medicine and materials processing. Real-time control comparisons with learning-based MPC with no uncertainty handling and offset-free MPC showcase the usefulness of LB-SMPC for predictive control of safety-critical systems with hard-to-model and/or time-varying dynamics.