The hippocampus is a world-model generator

Standard textbooks describe the hippocampus as a memory structure. Damage it, and the patient cannot form new long-term memories. This framing is correct and incomplete.

What the hippocampus actually does, according to a generation of research starting with O'Keefe and Nadel and culminating in the 2014 Nobel for grid cells, is build cognitive maps. Spatial maps for navigating environments. Conceptual maps for reasoning about non-spatial structures the same way. The "memory" function falls out of this: remembering is replaying a position in a built map.

This reframes a current AI debate. World models, the architectural bet that AI needs internal generative simulators, are usually pitched as a future research direction. The brain has had world models all along, housed in the hippocampus, doing exactly the work LeCun argues current AI is missing.

The interesting implication is not that AI should copy the hippocampus. It is that AI should expect the same architectural separation the brain found necessary. Pattern recognition can live in cortical layers; world-model generation may need its own structure, sized differently, integrated through bridging connections rather than fused. The labs trying to make one architecture do everything might be fighting a fight the brain already lost.