How do we continually adapt to novel situations while keeping past memories intact? Building on theoretical work by the Costa Group at DPAG (Pemberton et al. 2024), a new study reveals how the brain balances skill reuse with memory preservation. By jointly tracking activity in the premotor cortex and cerebellum, researchers monitored mice as they learned two distinct tasks built on the same core structure: perform an action, wait, and receive a reward.
In a new study just published in Nature (Garcia-Garcia et al. 2026) the researchers discovered a clear division of labour between the neocortex and the cerebellum. While the cortex reuses identical activity patterns across related tasks, the cerebellum takes a different approach: it reorganizes — or effectively ‘rotates’ — these shared patterns. This geometric rotation keeps distinct tasks from colliding while preserving their underlying structure, offering a new blueprint for how the brain balances generalization with task-specific precision.
Being able to reuse previous knowledge without different memories or skills interfering with one another is a fundamental problem for learning. The study suggests that the cortex and cerebellum may solve different parts of this problem: the cortex provides reusable building blocks for behaviour, while the cerebellum adapts these building blocks to the current situation.
The findings also challenge the traditional idea, first introduced by Marr (1969) and Albus (1971), that cerebellar granule cells mainly separate information by transforming it into increasingly complex, high-dimensional representations. Instead, the cerebellum may preserve useful structure while reorienting it to separate different contexts. This principle could ultimately help us understand how the brain learns many skills efficiently, and may also provide inspiration for artificial systems that need to learn multiple tasks without them interfering with one another — an open problem in artificial intelligence.
The study was a collaboration between the National Institute of Neurological Disorders and Stroke at the US National Institutes of Health (NIH) and DPAG. Experimental work was led by Martha Garcia-Garcia in the lab of Mark Wagner at the NIH. At Oxford, Michał Wójcik, a postdoctoral researcher in the Costa Group, led the computational modelling. By testing how different neural architectures support learning while preventing task interference, his simulations — paired with the experimental findings — led the team to challenge long-standing theories in the field while offering a compelling solution to the continually learning problem.