Artificial neural networks have been applied to engineering design retrieval. ART-1 networks are used to adaptively group together similar engineering or graphical designs. The information used to group the parts is coded into binary representations which, in their basic form, amount to bit maps of design descriptors. This technology has been used to build neural databases for the retrieval of two- and three-dimensional engineering designs. The authors discuss a feasibility-level system that learns to group sheet metal parts for modern airliners into similar families, and then to recall the best matching family when presented with a new design. An addition to the algorithmic form of ART-I was introduced that allows it to operate directly on runlength encoded vectors, and to generate compressed memory templates. When compared to the regular uncompressed algorithm on real engineering designs, the performance of this compressed algorithm demonstrated a significant savings in storage of the input vector and the memory templates.<<ETX>>