Encoding activity predicts distinct recognition signals during working and long-term memory in the human hippocampus

Neurocomputational models predict that episodic memories are represented by sparse, pattern-separated hippocampal codes. Recognition within working memory capacity may instead rely on overlapping cortical codes. We analyzed human intracranial microwire recordings during a continuous recognition task to test whether retrieval demand and spiking activity at encoding determine which neural signal is expressed. We show that relative increases in firing from pre-stimulus to post-stimulus onset at encoding predicted an item-specific signal that was present only in the hippocampus and only during long-term memory retrieval of remembered items. In contrast, consistently elevated firing predicted a generic signal that distinguished remembered from novel items, was present in the hippocampus during both working and long-term memory retrieval, and extended to the anterior cingulate and prefrontal cortices during working memory. These findings are consistent with a new theory that distinguishes working memory from long-term memory in terms of the neural coding scheme accessed at retrieval.