James Le describes video as a combination of visual events, speech, sound, text and motion unfolding over time. Isolated frame descriptions lose continuity, so useful memory must preserve entities, sequences and relationships across clips.
James Le presents an ingest-once, reason-many approach: semantic chunks and embeddings feed a context store, while retrieval and reasoning return evidence linked to relevant timestamps. The architecture separates durable video understanding from a particular question or agent task.
James Le connects that memory to an agent harness that plans retrieval, uses tools and constrains depth, scope, cost and output. Sports and traffic demonstrations illustrate possible workflows, but they do not establish general benchmark superiority or production accuracy.
Watch on YouTube
