Markdown source

Video Has No Memory. Here's How We Built One.

Conference Context

Session Description

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.

Media Evidence

No related AI Engineer channel video found yet.

Evidence Graph

This evidence graph is generated from currently linked source material: official schedule text, related video pages, cached transcripts, visible slide text, dense/reconstructed slide pages, and AI slide-classification audits.

Media Signals

No linked video, transcript, or slide source has been attached yet.

Agent Reading Notes

Use these signals to refine the synopsis, topic links, people/company context, and method notes. If a source is a related external video rather than an exact official recording, keep it framed as supporting evidence.

Transcript Status

No official session recording transcript was found by exact title match on the AI Engineer YouTube channel during this run.

People

Notes

Synthesis

Synthesized Breakdown

Video Has No Memory. Here's How We Built One. ## Conference Context - Date/time: 2026-07-01 · 2:25pm-2:45pm - Track/room: Graphs · Track 5 - Speaker(s): James Le - Session type/status: sponsor · confirmed - Track: Graphs - Room: Track 5 - Session type: sponsor - Status: confirmed ## Session Description 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.

Speaker And Company Context

Topics Covered

Derived Links And Source Material

Novel Concepts / Clever Methods

Evidence Boundary

This synthesis is based on the official schedule and linked source pages. It should be revisited when exact session recordings or transcript-backed secondary sources are available.