A team of neuroscientists at Columbia University has found that the brain does not store a given memory in the exact same neurons each time it recalls it. Instead, the neural pattern shifts over days and weeks, a phenomenon the researchers call "representative drift."
According to the study, the same odor prompted different groups of neurons to fire across multiple recording sessions, even though the animal could still tell the scents apart. The finding points to a fluid approach to memory that current artificial neural networks cannot match.
What the researchers actually saw
In the lab, mice were repeatedly exposed to specific odors while electrodes recorded activity in the piriform cortex, the brain region that processes smell. Performance of a classifier trained on day-one recordings dropped to near-random levels after 32 days. Fear conditioning did not lock the responses in place either; the firing pattern kept moving.
Daily re-exposure slowed the drift but did not stop it. Once exposure stopped, the rate increased again. So the brain was still recognizably tracking the same concept, just on shifting hardware.
| Condition | Drift Rate | Classifier Accuracy |
|---|---|---|
| First exposure | Baseline | High |
| Daily exposure | Slowed | Declining |
| Exposure stopped | Accelerated | Worsening |
| 32 days later | Significant | Near-random |
How this differs from the textbook story
Older neuroscience leaned on the idea of stable "grandmother cells"—small, fixed circuits that reliably fire for one concept every time. The Columbia work pushes back on that view: the brain keeps the memory, not the address.
Smell sits in a useful middle ground for this kind of work. It is more grounded than abstract thought ("what does London feel like?") and more flexible than the hardwired link between motor cortex and muscles. That makes it a clean place to watch memory reorganize itself.
Our analysis
Today's neural networks are essentially frozen between training runs. You teach them on a dataset, freeze the weights, and they produce answers from that fixed map. If biological memory works by continuously redistributing its internal code, then our AI systems may be missing something fundamental about integrating new experience without full retraining.
That does not mean tomorrow's chatbots will forget your preferences. It means the field is being pointed toward architectures that can re-route their own internal representations, the way a brain routes "cotton candy" through fresh neurons each time you smell it.
The analogy the researchers use is an FM radio in a car: the station stays the same, but the receiver keeps retuning to account for static, distance, and interference. The brain appears to do the same with memory—not deleting old associations, but constantly restoring its connection to reality while spreading data across a shifting network.
Who should care
- AI researchers**: The findings challenge the assumption that learned patterns must live at fixed addresses. Drift-like mechanisms could unlock true continual learning.
- Security and fraud teams**: Models trained on last quarter's attack patterns become obsolete fast. Architectures that redistribute learned features internally would adapt without expensive retraining.
- Neuroscientists**: The work opens questions about whether drift affects language, decision-making, and emotional memory—or if olfaction is uniquely dynamic.
- Skeptics**: This is one study in mice, focused on one sensory system. The leap to human cognition and production AI is speculative for now.
What to watch next
Keep an eye on whether other groups reproduce the drift finding in different brain regions, especially in areas tied to language and decision-making. If the effect is broad, expect a wave of papers adding drift-like mechanisms to artificial networks. If it stays isolated to olfaction, this remains an important but narrower piece of the memory puzzle.
The brain is messier and more adaptable than our best AI. A mouse study just gave researchers a new clue about where to look.
Key takeaways
- Mouse brains continuously reshuffle which neurons represent a given memory while keeping the memory itself intact.
- The drift speeds up or slows down based on how often the stimulus is encountered.
- Standard AI models freeze their internal patterns between training runs, a contrast that may explain their struggle with continual learning.
- Adopting drift-like mechanisms could help security and fraud-detection tools adapt as attacker behavior changes.
- The finding is limited to one sensory system in mice; broader reproduction is the next test.
Frequently asked questions
What did the Columbia University study actually discover?
It found that mice use different neuron groups to represent the same odor at different times, a process the team called representative drift. Memory stays intact even as the underlying neural pattern shifts.
Is this the same as forgetting?
No. The animal could still distinguish the scents. The memory persists; only its physical location in the brain moves around.
How does this relate to artificial intelligence?
Today's neural networks store learned patterns in fixed weights. Biological drift suggests a way to keep memories usable while allowing the internal code to reorganize, something current AI systems struggle to do.
Was this research done on humans?
No. The experiments used mice and focused on the olfactory system. Human memory is far more complex, and translating the findings is an open question.
Where can I read the original paper?
Search the term "representative drift piriform cortex" to locate the Columbia team's published study.