From moving a cursor to mending a memory
In Volumes I and II you learned to decode what the brain sends out — movement, speech, cursor intent. A cognitive prosthesis targets something categorically harder: the brain's internal operations. There is no muscle, no articulator, no obvious behavioural read-out. The signal you want — a forming memory, a shifting focus of attention — is distributed, content-specific, and largely private.
Why take on such a hard target? Because the clinical stakes are enormous. Amnesia from hippocampal damage, memory loss in dementia and traumatic brain injury, and disorders of attention and executive control are among the most disabling conditions in medicine — and among the least served by today's neuroprostheses, which overwhelmingly address the motor and sensory periphery rather than cognition itself.
What is a cognitive prosthesis?
A motor decoder observes neural activity and reads out an intended command. A cognitive prosthesis goes further: it participates in the computation. The canonical example is the hippocampal memory prosthesis, which does not merely record a memory — it replaces a lost neural transformation, taking the input a damaged region can no longer process and writing back the output that region would have produced.
The core idea in one line: learn a parametric map F_θ that approximates the transformation H a healthy circuit would have applied to its input x(t), then implement F_θ in silico and write its output back into the brain.
Notice what this requires that a cursor decoder does not: an accurate encoding model of the target circuit, a way to write patterned activity (not just read it), and confidence that the circuit's own downstream machinery will interpret the written output as legitimate neural code.
The read/write asymmetry, sharpened for memory
Across BCI there is a deep read/write asymmetry: we can record neural activity far better than we can inject meaningful activity. Memory sharpens this to a point. To write a memory you must know its code — and the neural code of a specific memory is exactly what we do not have.
This is why the honest frontier separates two problems. Reading cognitive state — is this person attending? encoding well? about to recall? — is progressing quickly. Writing specific content — installing or editing a particular memory — runs headlong into the biomimetic-write problem: an ill-posed inverse of an encoding model we understand only in part.
Demonstrated versus aspirational
Be precise about the state of the art. In small cohorts of epilepsy patients with implanted electrodes, closed-loop stimulation timed to memory-relevant brain states has produced modest but real improvements in recall — on the order of tens of percent on narrow laboratory tasks, measured within-subject. That is a genuine milestone, not a marketing claim.
Equally important is what has not been shown. No one has recorded a specific memory and replayed it into another context; no one has installed knowledge or uploaded a skill. General memory implants remain science fiction. The rest of this track walks the line between these two — real, narrow results, and the deep open problems beyond them.