We consider a dynamic reservoir of spiking neurons to generate routing decisions for bundle transmissions in a delay tolerant network (DTN). The reservoir transforms the context, as described by selected system metrics, and the temporal information that is contained in the network's contact list into a high-dimensional feature space, which allows quickly determining the next-hop for bundles whenever needed. The method can be of particular interest for space networks, which are characterized by resource-constrained nodes, known contact opportunities, and dynamic features, such as the state of the channels and buffers. The learning process occurs offline using linear regression and involves a modest computational effort. After training, the spiking neural network can be used to achieve an efficient network operation, allowing to substitute the costly per-bundle shortest path computation with a simple "recall" action of the best next-hop for the current conditions as stored in the reservoir. We investigate different methods to translate the contact list information into a suitable input for the reservoir and evaluate the approach considering the impact of the readout function size applied to GEO-LEO DTN scenarios.