Box has teamed up with Nvidia to give enterprise AI agents a better memory. At its BoxWorks 2026 developer summit on November 5, the cloud content company unveiled Box Context Retrieval, a new engine built on Nvidia's Nemotron Retriever models that helps AI agents find the right passages inside company files — without reading every document first. Nvidia chief executive Jensen Huang joined Box chief executive Aaron Levie on stage for the opening keynote to present the feature.
The announcement targets one of the most practical obstacles to deploying AI agents inside companies: an agent that needs a single fact from a contract or a slide deck currently has to wade through far more material than the job requires, burning tokens and risking exposure to files it should never see. Box says its new engine fixes that by turning enterprise content into a structured knowledge layer that agents can reason over.
What Context Retrieval does
Box Context Retrieval automatically builds and maintains a secure index over the content a company keeps in Box. It uses Nvidia's Nemotron 3.5 Retriever models — Embed, Rerank and Parse, released in October — to pre-process documents into semantically rich metadata. When an agent asks a question, the engine returns only the chunks that best match it, instead of handing over whole files.
The system is aimed squarely at unstructured data: the contracts, slide decks, PDFs and spreadsheets that make up the bulk of what companies store, along with scanned documents, images, audio and video. Box says every request still checks the requester's live Box permissions, so an agent can only surface files its user is allowed to open, and compliance and audit controls stay intact.
The numbers — and the caveats
Nvidia's vice president of generative AI software, Kari Briski, told TechCrunch the company expects queries to run four times faster, with 20 percent fewer errors and 90 percent lower token consumption than traditional retrieval methods. Box frames the gain as 20 percent better search recall. These are vendor projections, not independently verified results, and real-world gains will depend on how much content a customer holds and how its agents are built.
Box is also bringing Nvidia's open-weight Nemotron models into Box AI, starting with Nemotron 3 Ultra, which the company says is available now to Enterprise Advanced customers. Context Retrieval itself is scheduled to reach general availability for those customers in early 2027. Box also announced Box Architect, a tool for designing, testing and deploying agentic workflows in plain language, plus new agent security controls.
Why it matters
The move shows where the enterprise AI race is heading: the bottleneck is no longer raw model capability, but the plumbing that connects models to a company's own information. Box wants to be the trusted layer where enterprise content and AI agents meet, rather than a storage bucket that other companies' tools raid for files. With agents moving from demos into daily work, whoever controls fast, permission-aware retrieval controls a large slice of the enterprise AI stack — and Nvidia, whose open-weight Nemotron models topped the RTEB retrieval benchmark in July, is now embedded in one more piece of it.



