If every memory a person holds could be written to a drive, would the person on the drive still be them? The question used to belong to fiction. Brain-computer interface research has now advanced to the point where neural signals associated with memory can be read, modeled and, in limited experiments, used to improve recall. This explainer sets out what has actually been demonstrated and where the hard limits remain.
A prosthesis for memory
The most direct evidence comes from a collaboration between Wake Forest Baptist Medical Center and the University of Southern California. In a study published in the Journal of Neural Engineering in March 2018, researchers worked with epilepsy patients who already had electrodes implanted in the hippocampus, the brain region central to forming new memories. They recorded the patients’ neural firing patterns while the patients performed a memory task, built a mathematical model of the patterns associated with correct recall, and then delivered electrical stimulation that mimicked those patterns.
According to the researchers, the stimulated trials showed about a 35 to 37 percent improvement in short-term and long-term retention of visual information compared with trials without stimulation. The work grew out of two decades of research led by USC biomedical engineer Theodore Berger, whose team first modeled hippocampal signaling in rats and monkeys before moving to human patients.
The result is often cited as proof that memory can be digitized. That overstates it. What the study showed is that the hippocampus’s encoding activity follows patterns regular enough to be modeled, and that reinforcing those patterns helps the brain store information more reliably. The content of the memories stayed in the brain; the device acted as an aid to encoding, not as a storage medium.
The hardware race
Several groups are building the interfaces that would make finer-grained recording possible. Neuralink’s N1 implant uses 1,024 electrodes distributed across flexible threads inserted into the cortex, and the company’s stated near-term goal is to let people with paralysis control computers by thought. Researchers at Brown University have taken a different approach with “neurograins,” microchips roughly the size of a grain of salt that are designed to be scattered across the brain’s surface in large numbers and to record from many points at once.
Both designs face the same obstacles. Electrodes that sit in brain tissue for years tend to provoke immune responses and scarring that degrade their signals, which is why much of the current engineering effort goes into flexible materials and coatings that the body tolerates. On the computing side, neuromorphic chips, which process information in ways loosely modeled on neurons, are being explored as a means of handling the volume of data that dense recording would produce.
Where AI fits
Reading signals is only the first step. Turning streams of neural activity into something meaningful, identifying which patterns correspond to which perceptions or intentions and building models of how memory is organized, is a data problem of enormous scale, and it is where machine learning has become indispensable to neuroscience. Most contemporary brain-computer interface work pairs the implant with AI models trained to decode the user’s neural activity.
That still leaves a gap between decoding and storing. Current systems can recognize that a particular pattern is associated with, for example, moving a cursor or recalling an image. No system has recorded the content of a complex autobiographical memory from the brain and reproduced it elsewhere, and there is no agreed account of how such a memory is physically represented in the first place.
The limits that are not technical
Even a complete recording would raise a question the engineering cannot settle. Human memory is not an archive of facts but an experience bound up with emotion: the smell of tea that brings back a childhood home, a melody that recalls a first relationship. A machine might store the record of a moment without anything that corresponds to how it felt.
And if a full copy were ever possible, the familiar puzzle returns: which is the real person, the one in the body or the one on the drive? Science may well answer parts of the memory question in the coming decades. The ethical and philosophical parts are likely to remain open for considerably longer.
