On-line handwriting recognition is considered here as a problem of classification of temporal sequences of symbol components. A neural-fuzzy architecture that extends and enhances Fuzzy ARTMAP for the categorization of sequences is proposed in this paper. A new local distance measure permits an efficient processing of the spatial patterns produced by the STORE memory model. This neural network is extensively analyzed in both mathematical and experimental terms, leading to important suggestions for the engineering design of this biologically inspired model. Finally, experimental results of the global handwriting recognition architecture are presented.