The key finding
A 2024 review in Trends in Cognitive Sciences challenges a foundational assumption about how our brains store temporary information. For decades, neuroscientists believed working memory—our ability to hold a phone number in mind or remember where we parked—relies on stable, unchanging patterns of brain cell activity. However, recent brain recordings show that neural activity during working memory actually undergoes dynamic variations before stabilizing. Importantly, when researchers build artificial neural networks optimized specifically for working memory tasks, these systems naturally exhibit the same dynamic patterns seen in real brains, suggesting this fluctuating activity isn’t noise but a fundamental computational feature.
What the study looked like
This was a theoretical review paper that synthesized findings across three domains: direct neural recordings from animals and humans performing working memory tasks, classical computational models of memory (such as attractor networks that assume stable activity), and modern task-optimized artificial neural networks trained using machine learning techniques. The authors examined why classical models predict stable neural coding while both real brain data and optimized artificial networks show dynamic activity patterns during the maintenance period of working memory tasks—the seconds-long window when you’re actively holding information without new input. Rather than conducting new experiments, the researchers analyzed the mathematical and computational principles underlying different modeling approaches to identify why they produce divergent predictions about neural dynamics.
Why researchers think this happened
The key lies in how different models are constructed. Classical attractor network models were built with hand-crafted architectures designed to produce stable states—essentially, researchers built stability into the models from the start based on the assumption that stability was necessary. In contrast, task-optimized networks are trained through machine learning to simply perform working memory tasks well, without pre-imposed constraints about how they should maintain information. These optimized networks consistently develop dynamic coding strategies on their own. The researchers propose that dynamic coding may actually be computationally advantageous: changing patterns could allow the brain to represent information more flexibly, maintain multiple items simultaneously with less interference, or prepare for upcoming actions while still holding information. The dynamic activity observed in real brains may therefore reflect an optimal computational solution that classical models missed by assuming stability was required.
How to read this carefully
This is a theoretical review synthesizing existing evidence rather than presenting new experimental data. The proposal that dynamic coding is “fundamental” remains a hypothesis that requires further testing—we don’t yet have definitive proof that the brain requires dynamic activity for working memory. The fact that optimized artificial networks show dynamic patterns is suggestive but doesn’t conclusively prove the brain uses the same computational strategy for the same reasons. Additionally, some neural recordings do show relatively stable activity in certain brain regions or task conditions, so the picture may be more nuanced than “all dynamic” versus “all stable.” The field is still working to understand when and why different coding strategies emerge. Replication across different labs, species, and tasks will be important for validating this new framework.
What this means for everyday life
This research shifts how we understand a cognitive ability we use constantly—holding information temporarily in mind while cooking, having conversations, or solving problems. While this won’t immediately change practical memory strategies, it suggests our brains may be more sophisticated than previously thought, using dynamic flexibility rather than rigid stability to juggle information. For those interested in brain health or cognitive training, this highlights that even basic cognitive functions involve complex, evolving neural processes we’re still discovering. It’s a reminder that scientific understanding evolves: what seemed settled—stable memory patterns—turns out to be an open question. Given this emerging view, approaches to understanding or supporting working memory may eventually need to account for dynamic rather than static processes in how our brains temporarily store the world.