Uvi stores company context so agents stop rebuilding it on every query

An AI agent asked a question about a billing exception at a large company often does the same work twice: it reads old Slack threads, old Jira tickets and old pull requests to piece together what the company already decided, and the company pays for that work on every query. Uvi, a context layer for enterprise AI agents, raised that complaint on The New Stack and gave it a price. Co-founder and CEO Gilad Salinger said that an answer put together that way costs roughly 75 cents to $1.41 and can go over $2.00, and that the same answer with Uvi in front of the agent costs 10 to 16 cents.

What Uvi is

Uvi describes itself as "the context layer for enterprise AI agents". The company page says the product has three parts: an intent layer that reconstructs why past decisions were made, a semantic layer that maps a company's own words to canonical terms, and a knowledge graph that links the people, documents and dependencies behind a topic. The agent that calls Uvi gets what the company calls "the smallest correct slice" of the data it needs for that one question, rather than a general search over everything.

The in-house test, and what it is not

Salinger told The New Stack that Uvi ran its platform against more than 900 real employee questions, using the same frontier model on both sides of the test: once with the model doing its own search over Slack, Jira and GitHub, and once with the model asking Uvi for context. On open-ended questions where the first agent got stuck in search loops, Uvi says it was up to 46 times cheaper. On the whole set, Salinger said the first approach cost roughly eight times more and took more than twice as long.

The New Stack notes that the tests were run by Uvi, inside Uvi, and have not been independently verified. The 900 questions, the frontier model used and the breakdown by question type were not published. Treat the figures as the vendor's own.

Where it runs today

The post says Uvi is in production at Lemonade and monday.com. The product page lists Atlassian (Jira), Slack and codebase connections and names Cursor and Claude Code as agent interfaces that already call it. No pricing is listed on the public site. The homepage also carries two further vendor metrics, 97 of 100 head to head wins against an MCP-enabled GPT setup and 2x lower latency against a RAG baseline, which carry the same vendor-measured caveat as the cost figures above.

A team shipping an agent that answers questions from a company's own data has one more vendor telling it the retrieval layer is where the cost sits. The numbers in this piece are Uvi's and Uvi has not shown its working, but the comparison is the kind a team can run on its own stack: take the same frontier model, measure a hundred of the questions real employees already ask, and compare the cost of the agent searching from scratch against the cost of the agent reading from a stored, permissioned context store. If the gap is anywhere near eight times, the retrieval layer is a budget line, not a side project.

Source

Adrian Bridgwater, "Developers are paying frontier prices for amnesia": Uvi wants to stop the agent relearning brain drain, The New Stack. Uvi product page.