V7's Context Graph: does file-grounded memory actually change your agent loop?

What it is

V7 Go is an agentic document platform. Its new piece — a "Context Graph" — is a structured memory layer built from your uploaded files, meant to be queried by agents at runtime instead of stuffed into a prompt. V7 Go is an agentic platform to build mission critical workflows, and organize buried context into memory that agents can query and act on. Under the hood it uses OpenAI models, including GPT-5.6 variants and GPT-6 Astra for the harder graph traversals.

The next-work-session test

Concrete scenario: you're building an agent that answers questions across a 400-PDF due-diligence pack. Today your loop probably does chunked retrieval → rerank → stuff top-k into context → hope. The Context Graph pitch is that entities and relationships get extracted once, then the agent queries the graph like a small internal API — so a question like "which portfolio companies share a board member with the target" doesn't require re-reading every file each turn.

Does that change your next session? Only if your agent is already bottlenecked on multi-hop retrieval. If you're doing single-doc Q&A, it won't move the needle over a decent RAG setup.

Pricing

Not publicly listed as a flat number. V7 Go pricing is structured around a base platform fee that includes access to the V7 platform with AI agents, plus user licenses and data processing charges based on actual usage and use case requirements. Third-party writeups confirm the same shape — pricing is assembled from platform access, users, and document-processing volume, and V7 does not publish a standard monthly rate. Expect a sales call. Pricing: partially unverified — no self-serve tier is published on the V7 Go pricing page.

Separately, if you're modelling token cost on the underlying models, V7 has published claims worth checking against your own workload: V7 measured a 78% lower cost per document using GPT‑5.6 Luna versus GPT‑5.4 mini. That's V7's claim on V7's workload, not a benchmark you should port blind.

What we'd actually use it for

Narrower than the marketing. The realistic developer use case is offloading one specific thing from your agent loop: entity-and-relationship memory across a fixed document corpus where the same facts are queried by many agent runs. Think contract review pipelines, deal rooms, claims files, regulatory filings.

If you already have a working agent and you're tired of writing custom extractors + a graph store + a query layer yourself, V7 is buying you that stack rather than a new capability. The OpenAI writeup frames it as institutional memory; in practice it's a managed context graph with an agent runtime on top.

Limits

• It's a managed service, not an SDK primitive. Your agent loop calls out to V7 rather than owning the memory. That's fine for prod, awkward for evals — you can't easily diff two memory strategies locally.
• Benchmark numbers are vendor-reported. GPT-6 Astra reached 89% accuracy on the hardest graph-query benchmarks versus 78% for GPT-5.6 Sol — that's V7's internal graph-query test, not a public eval. Build your own eval set before committing.
• Memory ≠ correctness. Persistent memory across sessions has its own failure modes. OpenAI itself has flagged that if an AI silently memorizes a hidden command today, how can you stop it from betraying your trust tomorrow — prompt-injection into long-lived memory is a real threat model you now have to design for.
• Domain fit is narrow. V7 Go is pitched at finance, insurance, and real estate document workflows. If your agent is doing code, ops, or general web tasks, this is the wrong shape.
• Lock-in. Once your Context Graph is the source of truth for agent memory, migrating off it means rebuilding extraction pipelines.

Try it if

• You're shipping a document-heavy agent to production and the retrieval layer is your biggest quality problem.
• Your questions are multi-hop across entities, not single-doc lookups.
• You'd rather buy a managed graph + runtime than maintain one.
• You have a real eval harness ready to test it against your current RAG baseline.

Skip it if

• You're early — a plain vector store plus reranking will teach you more about your problem.
• Your agent loop is code, tool-use, or browsing, not documents.
• You need the memory layer to be inspectable and versioned inside your own repo.
• You can't get pricing clarity from sales before your pilot ends.