BCIwiki (bciwiki.com) — A team led by UCLA researchers has shown that a transformer neural network trained on single-neuron spikes recorded through intracranial microelectrodes can predict which specific people or places participants would name as they recalled an audiovisual episode, according to a preprint posted to bioRxiv on September 5, 2026. The paper, 'Reading specific memories from human neurons before and after sleep' (DOI 10.1101/2025.07.01.662486), has not been peer reviewed.
The models were trained on spikes recorded during a single viewing of the episode. Using cross-channel attention, they combined signals from recording sites across the brain to detect patterns spread over regions and timescales. Each participant had an individual model, which was then tested on out-of-distribution memory recall.
Decoding depended on sleep. Models trained only on medial temporal lobe spikes decoded concepts significantly before sleep but not after, while models trained only on frontal cortex spikes decoded concepts after sleep but not before. Better frontal cortex decoding was associated with more non-REM sleep. The authors say the results point to memory information distributed across neural populations and transformed over wake/sleep cycles.
The authors suggest that decoding internally generated memories could lead to brain-computer interfaces (BCIs) that treat episodic memory disorders by enhancing or muting specific memories. The paper reports decoding results only and does not test any intervention.