Research Suggests How Working Memory Retains, Manipulates, and Deletes Specific Items

Research Suggests How Working Memory Retains, Manipulates, and Deletes Specific Items

Posted: May 16, 2023
Research Suggests How Working Memory Retains, Manipulates, and Deletes Specific Items

Story highlights

Researchers propose that “spatial computing” explains the specificity and broad flexibility of short-term working memory operations—which are impaired in schizophrenia and bipolar disorder. They suggest millions of networked neurons act as “generalists,” able to manipulate items without being burdened by computations that would reveal their specific contents.


One of the remarkable capacities of the human brain is its ability to remember things that we experience, see, hear, feel, smell, learn about—even things that we think about or that we thought about a minute, a day or a decade ago.

Working memory is one of several kinds of memory that we depend upon each day. It is a form of short-term memory that comes into play when we set our house keys down in a particular spot, or are told the name of a new acquaintance, or need to remember where we parked the car.

Working memory is impaired in a number of psychiatric illnesses. It is a core cognitive impairment in schizophrenia and bipolar disorder. It can also be affected in people diagnosed with psychosis and ADHD, among others.

“Working memory is a mental sketchpad for the short-term storage and top-down control of information,” a research team notes at the beginning of a newly published paper appearing in Nature Communications. The element of control, the team says, is at the heart of working memory’s central role in cognition. “We can select what we retain” over the short-term, they note, and we can also “read out or delete items from working memory as well as manipulate its contents.”

Led by Mikael Lundqvist, Ph.D., a 2017 BBRF Young Investigator at MIT and the Karolinska Institute, Sweden, and senior team member Pawel Herman, Ph.D., of the KTH Royal Institute of Technology, Sweden, the team’s experiments led them to propose a new way to understand how working memory works in the brain’s prefrontal cortex.

Their insights build upon recent research suggesting the central importance of bursts of neural activity in the cortex at two frequencies, called gamma and beta. Gamma waves are generated by the activity of neurons that are firing rapidly; beta waves are generated by much slower neural oscillations. It has been theorized that the brain’s ability to encode and flexibly manipulate items in working memory are the result of interactions between neurons oscillating at these two frequencies, gamma and beta.

Although there is experimental evidence that supports this idea, one important question has to do with how working memory can act on specific items. Both gamma and beta oscillations involve the combined activity of millions of neurons acting in coordinated fashion. How can these millions of networked neurons be selective enough to control the contents of individual items in working memory, which is highly dynamic? For example, once you have been to the supermarket and purchased foods you had previously noted were in short supply, how does the brain delete this bit of working memory from the inventory of things retained, or demote it to a position of relative unimportance?

In their newly published paper, the researchers provide experimental evidence supporting the idea that “spatial computing” explains the specificity and broad flexibility of working memory operations. Key to this proposal is the concept that selective control of items held in working memory is the result of processes in which the brain uses “network space”—actual physical locations in vast neural networks in the prefrontal cortex that oscillate fast (gamma) and slow (beta).

The evidence comes from experiments conducted in five rhesus monkeys, which were fitted with electrodes that recorded neuronal oscillations in the prefrontal cortex. Recordings were made while the animals performed various working memory tasks. Some involved having to remember sequences of objects or colored squares; others involved matching related objects.

The team believes that representations of items in working memory are “moved across the spatial dimensions of the cortical network depending on task demands.” Assigning or moving the representation of an item from one part of the network to another can, for example, change the temporal order of the item, or whether it is currently a “priority” item or not.

“Control in working memory thus comes from where in network space a specific working memory item is held,” the team wrote. This allows it to be accessed and “operated upon” just by knowing its place in network space. This is crucial, the researchers believe, and amounts to a potential advantage for the properly functioning brain, since it enables control over working memory items without having to perform complex calculations that would reveal their specific identity or content, or what constellations of connected neurons form the actual memories of each item. Difficulties performing this spatial control operation in the cortex could perhaps be present in psychiatric disorders in which working memory is impaired.

Another way of thinking about the ability of the healthy brain to manipulate items in working memory is to appreciate its efficiency. It may be evolved to devote a minimum of the brain’s computing resources to important tasks. In this case, assuming the research by Drs. Lundqvist, Herman and colleagues is replicated, millions of networked neurons engaged in working memory tasks are, in essence, acting as “generalists,” able to manipulate items without being burdened by computations that would reveal their specific contents. The same networks could be used to manipulate multiple memory items.

An activity like working memory manipulation that is shared across many neurons and across contexts is called “low-dimensional activity” by neuroscientists. The team believes that because of the “generalizability” of working memory operations in the cortex, “spatial computing allows items that are novel to a specific task to be operated upon without having to re-train networks to [learn about] the new items.” Indeed, the team suggests, low-dimensional activity may eventually be shown to “dominate cortical activity” involved in cognitive operations performed on memory representations. Future research could clarify the implications of this insight for memory and other cognitive dysfunctions in psychiatric illnesses.

The research team also included Earl K. Miller, Ph.D., 2016 BBRF Goldman-Rakic Prize winner; and Melissa R. Warden, Ph.D., 2012 BBRF Young Investigator.