Multiple Routes to Metacognitive Judgments of Working Memory in the Macaque Prefrontal Cortex

The ability to evaluate one’s own memory is known as metamemory. Whether metamemory is inherent to memory strength or requires additional computation in the brain remains largely unknown. We investigated the metacognitive mechanism of working memory (WM) using two-photon calcium imaging in the prefrontal cortex of macaque monkeys, who were trained to memorize spatial sequences of varying difficulties. In some trials, after viewing the sequence, monkeys could opt out of retrieval for a smaller reward, reflecting their confidence in WM (meta-WM). We discovered that PFC neurons encoded WM strength by jointly representing the remembered locations through population coding and their associated uncertainties. This WM strength faithfully predicted the monkeys’ recall performance and opt-out decisions. In addition to memory strength, other factors— trial history and arousal—encoded in baseline activity predicted opt-out decisions, serving as cues for meta-WM. We identified a code of meta-WM itself that integrated WM strength and these cues. Importantly, WM strength, cues, and meta-WM were represented in different subspaces within the same PFC population. The dynamics and geometry of PFC activity implement metacognitive computations, integrating WM strength with cues into a meta-WM signal that guides behavior.