The DARE architecture is an emotion-based agent model. This model is based on a double-representation of stimuli: a complex representation, oriented towards recognition, and simple one, oriented towards feature extraction. These two representations are associated and stored in the agent memory. This paper describes a formalization of an indexing mechanism. Given a new stimulus, the goal of this mechanism consists of finding in memory, efficiently, using the simple representation, the best matching item according to the complex representation. To assess the efficiency gain of this mechanism, an implementation was devised using the classic recognition problem of handwritten digits.