The Brain Saves Memory Like a Cashier at Rush Hour express

Original: "Cortex and subcortex play distinct roles over learning when cortical memory is limited"
· Matthew Farrell, Taro Toyoizumi
arXiv:2606.00667 · 2026-05-30 · CC BY 4.0 · 1 min
The cortex learns rules, while deep regions chase rewards — together they work around limited memory.

📄 Showing the "Simple" version — "Popular" is not ready yet. Add it to favorites to help prioritize it.

A supermarket cashier at rush hour doesn't memorize every scanned item: they keep the running total in their head and watch the queue. The brain allocates its resources similarly: the expensive, precise cortex (responsible for deliberation) and the simpler, economical subcortical nuclei (where habits live) work in tandem. For a long time, it was thought that the cortex builds a map of the world, while the subcortex simply remembers where the reward is. But new work shows that when rewards change frequently, the cortex benefits from remembering not the rewards themselves but the general rules of the game.

In the experiment, a model with limited memory learned to choose between two doors behind which a treat would appear and disappear. If the treat moved often, the 'smart' module stopped remembering where it currently was and instead learned that the treat tends to move. It's like remembering not a friend's address but that they love moving every six months.

This approach offers a new explanation for an old puzzle: why the cortex and subcortex seem to speak different languages during learning. The cortex picks out the general structure—'there's a pattern here'—while the subcortex notes: 'this is where they give candy.' When the world is unstable, the memory-limited cortex stops chasing every candy and becomes a strategist that learns rules, not facts.

The memory limit in the model was strict — only a few 'slots' for storing information. That was enough to remember either the exact coordinates of the reward or the general rule. In an unstable environment, the rule won, even if it caused the model to make occasional mistakes in the moment.

In 'The Matrix,' people learned instantly by uploading programs directly into their brains. This work suggests that the cortex might not just memorize data but absorb structure, like loading not a file but an operating system.

Concepts
learning Reinforcement Learning neural network memory architecture working memory training information storage hippocampus game theory brain neural network decision making frame generalization training data visual cortex control theory agent model selection parameter reservoir computing