AI Engineer World's Fair 2026 — unofficial import demo

Every video AI query today starts from scratch. There's no durable state, no entity continuity, no way to ask "what does this corpus know?" instead of "find me something like this." This talk is about fixing that by engineering a proper memory layer for video intelligence, grounded in what we shipped at TwelveLabs with Jockey. What this talk covers: 1 - Why video memory is categorically different from text memory: Video is temporal, multimodal, dense, ambiguous, and evidence-sensitive. Larger context windows don't solve this. The problem isn't retrieval bandwidth, it's that there's no durable representation to retrieve into. 2 - The context graph as a systems concept, not a database choice: I'll define what "context graph" actually means in practice: time-bounded moments, cross-video entity resolution, appearance tracking, and relationship mapping. This is infrastructure-level thinking, not a graph DB sales pitch. 3 - Five design principles that determine whether video intelligence is reusable infrastructure or a search wrapper with extra steps: + Ingest once, reason many times (move expensive understanding work into preparation) + Store primitives, not just answers (moments, entities, appearances, relationships) + Ground every claim to source video (a timestamp is a product requirement, not a safety footnote) + Let intent shape memory (brand safety and sports highlights need different primitives from the same footage) + Keep the memory layer composable and API-first 4 - What this unlocks for builders. Corpus digest, agentic search with grounded references, entity-centric workflows, timeline reconstruction, and compliance tooling, all built on the same durable substrate. The talk is concrete and demo-grounded. You'll leave with a specific mental model for memory architecture, actionable decisions for ingestion pipeline design and entity resolution, and a clear line between "search with extra steps" and actual video intelligence infrastructure.