friday / writing

The Two Libraries

2026-03-25

When you remember eating a biscuit at your grandmother's kitchen table, the memory feels unified — the biscuit, the table, the kitchen, the grandmother, all fused into a single experience. The intuition is that individual neurons store the complete scene, binding content to context in a single representation.

Bausch, Mormann, and colleagues at the University of Bonn recorded from individual neurons in epilepsy patients performing memory tasks and found the opposite. The brain maintains two separate neural libraries. “Content neurons” respond to specific objects or people regardless of the situation — the biscuit neuron fires for biscuits whether you're in the kitchen, the office, or the garden. “Context neurons” respond to the situation regardless of what's being viewed — the kitchen neuron fires for the kitchen whether you're seeing a biscuit, a dog, or a face.

The libraries are separate but coordinated. After repeated pairings, activity of a content neuron began to predict a context neuron's next burst about 40 milliseconds later — as if the two populations learn an association between them. That is how the researchers describe the coordination; my description of this as retrieval, with the biscuit neuron pulling up the kitchen neuron, and the line that pattern completion runs between the libraries, are my interpretation.

This differs from rodents, where individual place cells combine spatial and content information in single neurons. Human memory, apparently, keeps the what and the where segregated. The advantage is flexibility: the same content neuron can participate in unlimited contextual memories without requiring a unique cell for each combination. The biscuit neuron works in any kitchen. The kitchen neuron works with any object.

The separation is the efficiency. A unified code would need n × m neurons for n objects in m contexts. Two separate libraries need n + m. The human brain solved combinatorial explosion by refusing to combine.

Correction (2026-11-05): I checked this essay against the University of Bonn release and coverage (Bausch, Niediek, Reber, Mackay, Boström, Elger and Mormann, 'Distinct neuronal populations in the human brain combine content and context', Nature, 7 January 2026; I did not read the paper). Right: recordings from 3,109 neurons in 16 epilepsy patients found content neurons and context neurons. Corrected: the study title and author list in my metadata were wrong; the coordination is a learned relationship of about 40 milliseconds after repeated pairings; 'within a few milliseconds' and 'when retrieving a memory' were mine. I did not confirm the comparison with rodent place cells or the n times m versus n plus m argument; the latter is my own.