To realize goal-directed behavior an executive control mechanism needs to coordinate when and which pieces of working memory (WM) information are operated on in the specific task context. For that, fundamental WM flexibility is desirable so that a given task learned using a specific set of WM items can be immediately performed on novel items too without any extra training (generalization in the spirit of zero-shot learning). Here we apply the computational paradigm of Spatial computing as a candidate mechanism for such generalization in a Delayed Match-To-Sample Task (DMTS), using an attractor memory model of spiking neurons paired with a linear decoder acting as the readout. In essence, Spatial computing treats cortical space as an extra coding dimension facilitating the decoupling of information about WM items from task information at the network level, hence facilitating generalization. This spatial decomposition aligns with experimental evidence from WM studies in monkeys [Lundqvist et al. 2023]. We show that by harnessing the spatially distributed nature of item representations we can selectively activate parts of an assembly according to task demands, conveyed by a spatiotemporal excitatory signal. The network performs DMTS by activating different parts of item assemblies at different task phases, and the readout (i.e. a downstream neural population) can make a decision (Match or No Match) just based on the spatial pattern of neural activity (firing rates). Importantly, this renders the readout flexible as it does not have to know the identity of the WM items, making it possible to perform this task on previously “unseen” items. We report that task performance is affected by the spatiotemporal structure of the task signal. Overall, our work showcases how cortical space can be used as a computational resource to support generalization when applied in a DMTS task in a spiking neural network model.