In this paper, a rapidly varying parameter estimation method for nonlinear systems is proposed. The parameters are supposed to be unknown functions of known causes, here referred to as associative memory variables. The estimation proposal consists of two stages; in the first one, an a priori estimation is calculated as the average of a dynamic subset of past estimates, defined by similarities between the current and past cause values. In this paper, this subset is referred to as associated memory. Each parameter of the model corresponds to an associated memory set, which is updated over time. In the second stage, the a priori estimation is updated using the innovation brought by the new incoming data and the recursive least squares algorithm. A nonlinear in the parameters moving average, a multi-parameter model, and a model with multi-dependent parameters are analyzed under this perspective. Two examples are shown to validate this proposal and alternative methods from the literature are used for performance comparison.