Latest
How to Give AI Agents a Memory of Your Company
Company knowledge is often scattered across files, folders, and systems. Using V7 Go’s Context Graph as a case study, we examine how that information can be turned into structured, source-linked memory that AI agents can use.
The problem agents keep hitting
Ask an AI agent to analyse a fund portfolio and it will begin by searching for context it almost certainly searched for last time: which report is current, how the same company is spelled in three different data rooms, which document supersedes which. Every request starts from scratch. This is not a model capability problem—it is a memory problem.
V7, a European startup founded in 2018 by Alberto Rizzoli and Simon Edwardsson after they built a computer vision accessibility app together, built its agentic platform V7 Go specifically around this gap. The pitch, as expressed on OpenAI's published case study page, is direct: "To solve hard enterprise use cases across finance and insurance, AI needs to learn how your business operates just as well as it learned from the Internet."
What the Context Graph actually does
When V7 Go connects to a repository—SharePoint, Google Drive or similar—it doesn't simply index text for keyword search. It scans for entities (companies, funds, people, any element defined in an ontology), identifies relationships and attributes, and writes the results into what V7 calls a Context Graph: a structured record where every fact is linked back to its original source document.
The practical consequence is that an agent querying the graph does not need to re-read thousands of files to know that Fund A's Q3 report supersedes Q2, or that "Meridian Holdings" and "Meridian Holdings Ltd" are the same entity. The graph is designed to preserve those relationships so the agent can query them without reconstructing the same context from scratch. When new files arrive, V7 Go updates the relevant records rather than rebuilding from scratch. If the graph doesn't contain enough to answer a query, the system falls back to retrieval-augmented generation (RAG) over the underlying documents—so the structured layer augments rather than replaces conventional search.
V7 describes this approach as "an order of magnitude cheaper and faster to traverse than long-context approaches," though that comparison is a vendor claim and Chooscope has not run independent tests.
Which models do what
V7 Go uses a division of labour across OpenAI's model family:
- GPT-5.6 Luna handles extraction and organisation—pulling information from millions of files and building the Context Graph. V7 reports 78% lower cost per document and an 11.6 percentage-point accuracy gain with Luna compared to a prior baseline.
- GPT-5.6 Terra and Sol handle reasoning and tool use inside multi-step workflows.
- GPT-6 Astra is used on the most demanding graph queries, including financial analysis across thousands of documents, where V7 reports 89% accuracy on its hardest query tests.
These figures come from V7's own reporting on the OpenAI case study page; they are not independently verified results. The baseline each number is measured against, and the exact test conditions, are not fully described in the available material.
The HERB benchmark result
V7 has also evaluated its retrieval system on HERB, a benchmark designed to test finding and connecting information spread across enterprise systems. V7 says its retrieval-only system outperformed the official HERB baseline by 69% and reduced hallucinations on unanswerable queries by 38%.
HERB is a specific, bounded test; it does not automatically represent performance across all document types or enterprise setups. What it does illustrate is the core design argument: that resolving entity relationships before an agent queries them reduces both missed answers and fabricated ones.
What this means for teams designing agent workflows
The design pattern V7 embodies—extract entities and relationships into a graph at ingestion time, then let agents query structure rather than raw text—is a practical response to a genuine bottleneck. Finance, insurance and real estate teams working with dense, version-controlled document sets are the stated target, because in those domains retrieval errors are not acceptable.
V7 Go claims agents can complete 50–100 step workflows in minutes at 99.9% accuracy, maintaining an auditable trail of every decision. That is a strong claim for production contexts; any team evaluating the platform would want to test it against their own document corpus and ontology before treating it as a given.
The more durable takeaway is the distinction the design forces: the difference between an agent that searches every time and one that queries a maintained record. The first grows more expensive as the document base grows; the second shifts more of the work to ingestion and aims to reduce repeated search costs at query time. Whether V7 Go delivers that in practice for a specific organisation depends on how well its entity extraction handles that organisation's particular vocabulary and file formats—two specifics the current public documentation does not fully detail.
Sources
Change history
- 承認mail が承認
- 無効新しい承認で置き換え
- rev.3Published
