Bio Mirror / Memory, Motor Control, and Embodiment
Working Memory and Its Hard Limit
A small, actively maintained buffer. The limit is severe, and it may be a feature.
The structure
Working memory is supported by sustained activity in prefrontal and parietal cortex, maintained by recurrent excitation and shaped by dopaminergic gating that determines what gets in and what is protected from interference. Recent evidence suggests some content is held in rapid synaptic changes rather than in persistent firing — "activity-silent" working memory.
The function it dictates
Working memory holds a small amount of information in an actively accessible state for immediate use. The capacity limit is famously small — around four items, not the seven of popular retelling — and it is one of the strongest constraints on human cognition, predicting performance across reasoning tasks. Content must be actively maintained and decays or is overwritten by interference.
The mirror in AI
A recurrent network's hidden state is a maintained buffer with a fixed capacity, and its failure mode — earlier content overwritten by later — is recognizably similar. A transformer context window is a different thing: large, passive, and not maintained by any active process. But the "lost in the middle" effect, where information in the center of a long context is used less reliably, shows that access within it is not uniform either.
The mirror in robotics
The state buffer in a behavior tree or task planner: the small set of facts about the current situation that the controller keeps live. Robotics engineers deliberately keep it small, for exactly the reason biology has to — reasoning cost grows with the state you must consider.
A transformer has no capacity limit of the biological kind; it has a context window, which is a different constraint with different failure modes. It is worth asking whether the human limit is purely a cost or partly a benefit — a severe capacity constraint forces abstraction and chunking, and a system that can hold everything has less pressure to compress. That question is open and it is a genuinely good one to argue about.
Open questions
- Is the four-item limit a constraint the brain works around, or a pressure that produces abstraction?
- Does the lost-in-the-middle effect have any principled relationship to biological capacity limits?
Related DEM-X entries
GI-DRFT-01Further reading
- Cowan, "The magical number 4 in short-term memory" (2001)
- Stokes, "Activity-silent working memory" (2015)