Recurrent networks with Generalising Boolean nodes, which are also known as GNUs (Aleksander, 1990a), are studied in this work in terms of their retrievability and attractor properties in the state space. Figure 1 shows a Generalising Boolean node with 4 inputs that has learned to fire 0 when the pattern at the inputs ABCD is more similar to 0000 and to fire 1 when more similar to 1111. The generalization in this case is defined by the similarity (measured here by Hamming Distance) between an unknown input and the two known ones. The Boolean function performed is seen in Figure 1.(a) in a Karnaugh Map form. It can also be seen that some entries are equally similar to both stored patterns (marked with? in the map) which can put the node under hesitation. Such “don’t care” entries are randomized 0 or 1 at each slot of time. Therefore, the node performs a different Boolean function at each slot of time and can fire different outputs for the same undefined input Such Boolean nodes are known as G-RAMs (Generalising Random Access Memories) as one of the ways found to implement them was to use conventional Random Access Memories and by randomizing the undefined entries in execution time (Aleksander, 1990b). Some work on storage capacity and temporal properties of such networks was done before (Wong and Sherrington, (1992), Braga, (1993a) and Braga (1993b)).