“Thoughts” can be likened to massively parallel webs of self-sustaining activation among the neural circuits of the brain.
One characteristic of the brain, which enables rich across-domain representations, is the ability to arbitrarily link the results of different neuronal sub-processes. For example, when we think of “an orange” we combine at least visual, auditory, and soma-sensory outputs from different cognitive levels. Even though the sub-processes are tailored by environment over time, the general blueprints are driven by our genes.
To allow us to develop machines that have even a hint of this capability, we need to allow access to the outputs of different sub-processes. Ignoring the fact that these sub-processes develop dynamically, would a shared memory space allow some of this cross-domain linking? For example, in an n-dimensional memory space, could a process arbitrary link to an output space of any other process? How would this linking work while allowing change in the sub-processes and the size of the output space?