Berkeley's StreamEmbed indexes your screen history on-device, 3.7x faster at matched accuracy
berkeley_ai · x · 2026-11-19
Researchers from UC Berkeley, Seoul National University and KAIST released StreamEmbed, an open-source system that embeds a screenshot every 5 seconds into a local index so you can search your past screen history in natural language, fully on a laptop.
- The latency-accuracy trade-off: OCR text indexes are fast but miss pixel-only info; full screenshot encoding is accurate but takes 2.3s per screenshot on a laptop GPU.
- StreamEmbed exploits redundancy in screenshot streams: spatial token pruning drops blocks repeating a neighbor, temporal KV reuse recomputes only what changed (57-71% less compute), plus a tiny adapter keeps embeddings faithful to the full screenshot.
- Results: 0.6-0.8s per screenshot, up to 3.7x faster than full-screenshot encoding at matched accuracy, +6.3 Recall@10 over OCR. Ships with an MCP server, Apache-2.0 licensed.
Related event: Berkeley-Led Team Open-Sources StreamEmbed for Screen History Search(3 posts)→
More from coding & agent
- GPT 6.2 Astra builds interactive 3D web app of Singapore Chinatown in one shot — OpenAIDevs · 2026-11-19
- Dev predicts 2027 will be the year of 'outcome engineering' for agents — hwbhatti · 2026-11-19
- Scenario open-sources a full AI art team as a Photoshop plugin, MCP-ready — aziz4ai · 2026-11-19
- Harrison Chase: the hardest part of agents is context — build a knowledge factory — hwchase17 · 2026-11-19
- How 30+ Silicon Valley Companies Are Competing for Programmers' Training Data — FinanceYF5 · 2026-11-19
- Kimi K3.1 Pro beats K3 on 14 of 17 benchmarks; pairing with Opus 5.5 halves spend — Kimi_Moonshot · 2026-11-19