This paper proposes an immunity-based RBF training algorithm for nonlinear dynamic problems. Exploiting the locally-tuned structure of RBF network through immunological metaphor, a two-stage learning technique is proposed to configure RBF centers and widths in the hidden layer. Inspired by affinity maturation process of immune response, immune evolutionary mechanism (IEM) with memory operations is implemented in the learning stages to dynamically fine-tune the network performance. Experiment results also demonstrate that the algorithm has reached good performance with relatively low computational efforts in dynamic environments.