OpenAI explains V7 Go’s source-linked memory for enterprise agents

OpenAI detailed V7 Go’s architecture for turning company files into source-linked context that AI agents can query during long-running enterprise workflows. The deployment routes work among GPT-5.6 models and exposes graph queries, ingestion and workflow operations to compatible clients through the Model Context Protocol, or MCP.

Structured extraction and other high-volume work go to GPT-5.6 Luna, while GPT-5.6 Terra or Sol handle chat, the Go Agent path, reasoning and tool use, according to OpenAI.

The deployment centers on V7’s Context Graph, which extracts information from company files and connects entities, relationships, facts, attributes and metrics to cited evidence. OpenAI says V7 Go can ingest repositories including SharePoint and Google Drive, attach extracted facts to new or existing ontology records, preserve citations to the original material and fall back to document retrieval-augmented generation when the graph lacks enough information.

For longer tasks, recent exchanges stay in the model’s active context while older material is stored in the graph for later retrieval. OpenAI says V7 Go can coordinate deterministic code, handoffs to smaller models, file generation and integrations while retaining an auditable trace of each run. The design is intended to keep organizational context available across a workflow without repeatedly reconstructing it from the underlying files.

The V7 MCP server exposes Context Graph querying, ingestion and workflow operations to compatible AI clients, with OAuth used for workspace authorization. OpenAI says customers can query and add material to the graph from ChatGPT and other compatible clients, and can create V7 Go workflows through MCP in Codex. A study of public MCP tools provides separate context on the broader protocol ecosystem.

V7 also supplied performance figures that are not independently established in the opened sources. The company says its agents can finish workflows spanning 50 to 100 steps in minutes with 99.9% accuracy, and that graph traversal is an order of magnitude cheaper and faster than long-context approaches. The public profile does not define the accuracy measure or disclose sample sizes, comparison systems, workloads or measurement methods for those claims.

In V7’s internal testing, the company says GPT-5.6 Sol reduced Context Graph tool-call errors from 2.7% with GPT-5.5 to 0.2%, selected workflows ran as much as 50% faster, and GPT-5.6 Luna lowered per-document cost by 78% compared with GPT-5.4 mini. V7 also says moving some PDF-heavy work from the Chat Completions API to the Responses API reduced token use by about 5% and improved caching reliability. The profile does not provide the document mix, token counts, run counts or full model settings needed to reproduce those results.

V7 reports that its retrieval-only system beat the official baseline on the HERB enterprise-search benchmark by 69% and cut hallucinations on unanswerable questions by 38%. The HERB paper describes a synthetic enterprise environment containing 39,190 artifacts and reports a best agentic-RAG average score of 32.96, with retrieval identified as the main bottleneck. It does not include V7’s results, so those comparisons remain company claims.

On a separate, harder graph-query set created by V7, the company reports 78% accuracy for GPT-5.6 Sol and 89% for GPT-6 Astra on the very-hard tier. The dataset and evaluation harness are not public in the opened sources.

The profile also attributes customer outcomes to the system, including 21-times-faster deal screening, a reduction in one review process from more than 100 hours to under 10 hours with $12,000 saved per task, and a 13.5% reduction in insurance claims-processing errors. The customers, cohort sizes, measurement periods and methods behind those figures are not disclosed. V7 says proactive triggers based on graph changes, inconsistency warnings and notices that an analysis relies on superseded facts are longer-term work rather than currently documented features.