TwelveLabs shows a video memory layer that stores moments, entities and themes
AI Engineer · youtube · 2026-07-23
In a talk titled Video Has No Memory. Here's How We Built One., TwelveLabs’ James Le argues that most video AI systems answer each query from scratch because video lacks a durable memory representation.
He says bigger context windows do not solve the core problem. Instead, the right approach is to treat video as a spatial-temporal volume and build a memory layer on top of it:
- ingest once, reason many times
- store primitives, not final answers
- ground every claim to timestamps
- let intent decide what gets remembered
TwelveLabs’ stack includes an embedding encoder, a context store, and a video language model exposed as an API. The demo examples span World Cup highlights, tracking Messi across a corpus, traffic security, and ad placement.
More from Multimodal
- Magnific teases October 28 release: 'We're making it move' — cuenca · 2026-10-27
- Midjourney --sref demo: retro Americana Batman art series — gizakdag · 2026-10-27
- AI video is a new medium, not cinema or animation — and it's perfect for depicting dreams — Kyrannio · 2026-10-27
- Dad turns his toddler's 31-year-old quote into a song via SUNO kids' choir — NathanpmYoung · 2026-10-27
- Band's local-AI music video hits 200k views: 900 generations kept 64, 3-second clips beat face drift — Hwang0500 · 2026-10-27
- Adobe Firefly video editor walkthrough: one-sentence Quick Cut turns 32s into 20s — LudovicCreator · 2026-10-26