You are holding a cup of coffee. Your eyes are closed. Your fingertip reads the curve of the ceramic and your brain — before you have consciously thought anything — already knows the shape and the object. It predicted it. The neocortex is the part of the brain that enables this prediction. It is a relatively recent-relative to human evolution, evolutionary development, and it is responsible for higher-order functions such as sensory perception, cognition, generation of motor commands, spatial reasoning, and language. Is there a possibility that the same predictive machinery that enables your brain to model the shape of a cup also enables it to model the shape of a social group? Does this have any bearing on how LLM's are architected for short and long-term memory?

The neocortical region: the "new brain" - is arranged in layers and columns. Connections between layers run mostly perpendicular to the surface, suggesting a columnar functional organisation. Some layers carry long-range excitatory lateral connections, enabling interactions between columns. Each of these structures is known as a cortical column. Every cortical column is further subdivided into mini-columns, with approximately 150–250 mini-columns per cortical column. A sensory input, say, a fingertip touching the surface of a cup with eyes closed - drives a prediction model inside the brain that is constantly evaluating the location of touch against the object being predicted. This is achieved by the firing of one or more mini-columns within a cortical column.

Now consider these in parallel. A widely cited, though still debated, finding in neuroscience proposes that the number of stable social relationships a human can maintain is determined by the relative size of the neocortex. This is measured as the ratio of neocortex volume to the volume of the rest of the brain. This ratio when mapped across the primate species is approximately 150, the Dunbar Number, proposed by Prof. Robin Dunbar through decades of research into the behavioural, cognitive, and neuroendocrinological mechanisms underpinning social bonding. 150 is not a flat threshold, it is nested. The layers are 5 → 15 → 50 → 150, each outer ring roughly three times larger than the one inside it, with emotional intensity decreasing as the group expands. This is because maintaining a cohesive relationship requires an individual to track not only their own state but coordinate continuously with every other member of the group; and that coordination carries a cognitive cost per relationship.

The threshold of 150 appears repeatedly in human organisational history. Dunbar points to the Roman maniple; the tactical unit of the pre-Marian legion, which comprised approximately 120–130 men, sitting squarely inside the 150 limit. Modern manufacturing companies have independently converged on the same constraint, capping production units and reporting structures below 150 while preserving the nested layers of 5, 15, and 50 within them.

The connection between these two observations runs deeper than analogy. The neocortex is, fundamentally, a prediction machine. Cortical columns do not passively receive input; they are actively modeling our world, running continuous predictions against incoming sensory data and updating when those predictions fail. Dunbar's research revealed that this same predictive machinery operates on social objects as it does on physical ones. Maintaining a stable social relationship is a computational act: our brain must hold a running model of another person — their history, intentions, likely behaviour — and keep that model current. Each relationship is, in effect, a persistent simulation.

The Dunbar number, then, is not a social threshold. It is a computational budget. The limit of approximately 150 stable relationships is the point at which the cost of running concurrent simulations of other minds saturates the available neocortical resource — the same resource that fires mini-columns to predict the location of touch against the object being held. The nested rings of 5, 15, 50, and 150 are not cultural conventions. They are the brain's own scheduling architecture: tighter rings receive more frequent model updates and carry higher fidelity; outer rings are maintained at lower resolution to preserve capacity for the core.

The question is whether this constraint is biological — or whether it is architectural, and therefore reproducible in any system that must maintain persistent, concurrent, contextually-dependent states.

In the age of large language models, the Dunbar number is not replaced; it is re-expressed. The constraint now lives in the context window, the coordination failure rate across agents, and the token budget. Efficient access to contextual information is tiered across short-term conversational memory, long-term knowledge stores, RAG pipelines, and LLM context windows — producing the same nested architecture the brain arrived at, for the same underlying reason.

References:

Lindenfors P, Wartel A, Lind J. 'Dunbar's number' deconstructed. Biol Lett. 2021 May;17(5):20210158. doi: 10.1098/rsbl.2021.0158. Epub 2021 May 5. PMID: 33947220; PMCID: PMC8103230.

Hawkins J, Ahmad S, Cui Y. A Theory of How Columns in the Neocortex Enable Learning the Structure of the World. Front Neural Circuits. 2017 Oct 25;11:81. doi: 10.3389/fncir.2017.00081. PMID: 29118696; PMCID: PMC5661005.