Abstract
The mechanisms underlying fast electrical signaling in neurons are well-studied. Long-term working memory, essential for cognition, is poorly understood. Imagine listening to an hour lecture, during which the speaker presents a sequence of words and slides that tell a story with many details. At the end of the talk, you ask a question that requires integrating these details with previous knowledge; this requires temporal context – linking words and images across time. The same cognitive functions occur during reading, as a sequence of fixations takes in groups of words whose meaning depends on their temporal context over many sentences and paragraphs. We have no explanation based on neural mechanisms for long-term working memory. How do brains encode temporal context over hours? How is long-term working memory used to generate cognitive behaviors such as thinking and planning? We can ask the same questions about transformers in AI, which might provide clues to long-term working memory in brains.