V7 AI Agent Memory: Context Graph and GPT-6 Results

V7 AI agent memory is an enterprise workflow design that turns company files into a source-linked Context Graph, then lets agents query that graph and fall back to document search with retrieval-augmented generation (RAG) when the graph lacks enough information. In an OpenAI customer story published September 21, 2026, V7 says it uses GPT-5.6 Luna for extraction, GPT-5.6 Terra and Sol for reasoning and tool use, and is starting to use GPT-6 Astra for its hardest graph queries. The reported results come from V7’s own systems and test harnesses; they are not independent benchmarks of every enterprise workload. Read the official V7 case study.

V7 AI agent memory: what the case study reports

The architecture addresses a practical limitation of document agents: relevant business facts can be spread across files, spreadsheets, data rooms, email and internal systems, while names and relationships vary between sources. V7 Go connects to repositories including SharePoint and Google Drive, identifies entities, facts, attributes, metrics and relationships, and stores those links in its Context Graph. Each fact can retain cited evidence back to the original source.

V7 describes this graph as a structured, current record that an agent can query rather than rediscovering the same relationships for every request. For a long-running agent, recent exchanges can remain in active model context while older material is retrieved from the graph when needed. That is the basis for calling the system institutional memory: the memory is not a claim that the model itself permanently remembers an organization, but a retrieval layer that the workflow maintains outside the model.

The affected users in the case study are primarily document-heavy teams in finance, insurance and real estate. V7 says its agents can complete 50-to-100-step workflows in minutes with an auditable decision trail. It also reports 99.9% accuracy for those workflows, but the public case study does not provide a full dataset, sampling method or independent reproduction. Treat the figure as a company-reported outcome, not a universal performance promise.

How the V7 Context Graph differs from document-only RAG

The official description does not present context graph vs RAG as an either-or choice. The graph is used to preserve resolved entities and relationships for direct querying. If it does not contain enough information, V7 Go can still search the underlying documents with RAG. In other words, the graph supplies reusable structure, while document retrieval remains a fallback for evidence that has not been organized or is better recovered from the source text.

V7 reports that its retrieval-only system beat the official baseline on HERB, an enterprise information-retrieval benchmark, by 69% and reduced hallucinations on unanswerable queries by 38%. Those figures describe V7’s implementation on that benchmark. They do not prove that a context graph will outperform every RAG system, or that the same gains will transfer to another company’s documents without equivalent ingestion, entity resolution, permissions and evaluation.

The customer outcomes are also specific to V7’s deployments. The company reports asset managers screening deals 21 times faster, reducing a full-day process to 15 minutes; one financial-services team reducing a review from more than 100 hours to under 10 and saving $12,000 in expert costs per task; and insurance teams reducing claims-processing errors by 13.5% against a manual baseline. These numbers are useful evaluation targets, but the article does not disclose enough detail to reproduce them from the public page alone.

Where GPT-5.6 and GPT-6 Astra fit

V7 assigns models by workload rather than sending every step to the largest model. GPT-5.6 Luna handles structured extraction and other high-volume work. GPT-5.6 Terra and Sol handle chat, the Go Agent path, multi-step reasoning and tool use. V7 says GPT-5.6 Sol reduced its Context Graph tool-call error rate from 2.7% with GPT-5.5 to 0.2%. It also reports GPT-5.6 Luna costing 78% less per document than GPT-5.4 mini in its measurements.

For its most difficult graph questions, V7 built a harder test set using messy data across thousands of documents. The company reports GPT-5.6 Sol scoring 78% on the very-hard level and GPT-6 Astra scoring 89%, while both were close to 100% on easy, medium and hard levels. This is a V7 benchmark, not an OpenAI-wide model score. Readers comparing it with another customer test can see our Perplexity GPT-6 Astra report; the two companies used different tasks and their percentages should not be compared as if they shared one test set.

V7 also says moving document-heavy workloads from Chat Completions to the Responses API reduced token use by roughly 5% for some PDF-heavy workflows and improved caching reliability. Again, “some” matters: the case study does not establish a guaranteed token reduction for every API migration. Our API overview covers the site’s broader API context without turning this customer result into a platform-wide promise.

Bringing the Context Graph into ChatGPT and Codex

V7 Go exposes Context Graph querying and ingestion through an MCP server. According to the case study, customers can use it from ChatGPT and other compatible clients, while Codex can create V7 Go workflows through MCP. V7 reports that this access, combined with simpler workflow design, reduced creation time for a medium-length workflow from about one hour to about 20 minutes.

That statement describes V7’s product integration, not automatic access for every ChatGPT or Codex account. The public story does not define pricing, plan eligibility, workspace approval, repository permissions or deployment steps. Teams should confirm those details with V7 and their own administrators before treating the connection as production-ready. For model and product entry points that OpenAI has separately documented, use our GPT-6 access guide.

A practical enterprise AI agent memory evaluation

The following checklist is our editorial recommendation, not an OpenAI or V7 certification and not a test we performed on V7 Go:

1. Choose representative questions that require facts from several documents, including renamed entities and conflicting versions.
2. Require every material answer to link to the exact source evidence, then inspect whether the citation supports the claim.
3. Include deliberately unanswerable questions and measure whether the system declines instead of filling gaps.
4. Track ingestion freshness: change one source fact, then verify when the graph and downstream answers reflect it.
5. Separate retrieval accuracy, reasoning accuracy and tool-call reliability so one strong aggregate score cannot hide a weak stage.
6. Measure latency and cost by document and by completed workflow, including retries and human review.
7. Test repository and workspace permissions with accounts that should and should not see the same source.

The core lesson is narrower than “graphs replace RAG” or “a stronger model solves memory.” V7’s published design combines maintained structure, source evidence, fallback retrieval, model routing, tools and an audit trail. Its reported GPT-6 Astra result suggests value on one demanding graph-query set, while the case study’s limits make a workload-specific evaluation essential. For another example of an agent company testing Astra on a different workflow, see our Cognition and Devin report.