Beyond motor intent
A passive BCI does not decode a command you deliberately send; it reads the mental state you are simply in — engaged or fatigued, focused or distracted, encoding well or poorly. This implicit read-out is the foundation of cognitive interfacing.
The applied face of this is neuroergonomics and the attention / workload BCI: estimating cognitive load or vigilance to adapt a task, a cockpit, or an interface in real time. Crucially, this is cognitive-state decoding — a read channel that needs no overt behaviour at all.
Encoding state and the subsequent-memory effect
A robust finding anchors memory read-out: neural activity during encoding predicts whether an item will later be recalled — the subsequent-memory effect. If you can detect a poor encoding state as it happens, you can intervene before the memory is lost.
A subsequent-memory classifier: from multi-region spectral features x(t) sampled at encoding, a regularised logistic model estimates the probability an item will be recalled. A low predicted probability flags a poor encoding state.
This classifier is exactly what drives encoding-state-triggered stimulation: in human studies, stimulation delivered during predicted poor encoding states improved later recall, whereas the same stimulation during good states could impair it. The state estimate — not the stimulus alone — carries the benefit.
Decoding intended recall
A more speculative frontier is decoding intended recall: reading out what a person is trying to retrieve, or which of several memories is being reactivated. Category-level and even item-level reactivation can be decoded above chance from intracranial signals — but this is a long way from reading arbitrary memory content.
Closing the loop for enhancement
Put the read and write sides together and you get brain-state-dependent, biomarker-driven enhancement: sense a cognitive state, decide whether to act, and stimulate only when it helps. This is the same adaptive logic as responsive neurostimulation and closed-loop DBS, turned toward cognition instead of seizures or tremor.
The sharpest lesson from these studies is that stimulation is state-dependent: the very same pulse can help or harm depending on the brain state it lands in. A cognitive prosthesis is therefore only as good as its real-time read-out of the state it is trying to improve.