The global workspace must consist of a population of neurons that are capable of holding general-purpose information. They are able to apply rules and operations on this information in a content-independent way. But all current neural implementations of this are content-dependent. In other words, if I were to learn a new kind of object, then all the operations that can be performed on objects would need to be updated to accommodate this new type of object. We have recently studied a neural mechanism that allows content-general ‘slot-like’ neural representations which allow the structural rules and operations to be encoded in separate populations of neurons to the content. We achieve this using rapid associative plasticity, so that the activity of a neural population no longer represents a constant thing. The pattern of neural activity corresponds to different things at different times. The “readout” of this variable pattern is achieved by retaining short-term information in synapses. In other words, the structure-content separation becomes a critical feature of synaptic short-term memory. Ultimately this factorisation allows neurons to hold and operate on information in a symbol-like way.