It is demonstrated how simulated annealing can be used in CBM (classifier-based model) associative retrieval to overcome the problems associated with gradient descent. It is shown that, with its use, the CBM can find a solution as required by the associative recall problem even in extremely disadvantageous object-space topologies where gradient descent fails. Simulated annealing is more robust than conventional gradient descent in reaching a globally optimal solution, as the results are independent of the initial placement of the cue. The use of fixed-temperature finite-length transition chains is shown to yield faster convergence than that of one inhomogeneous temperature chain, as used by H. Szu (1987). A modification of D.S. Johnson's (1986) formula for determining the initial temperature is likewise found to produce improved results. Appropriate terminating conditions are determined empirically, and solutions are shown to be within an acceptable accuracy level