How V7 gives AI agents institutional memory

Models can reason through complex tasks, but they don’t automatically know which fund report is current or how the same entity is named across three systems. That context lives in documents, data rooms, spreadsheets, and tools — scattered and invisible to agents. For finance, insurance, and real estate, retrieval mistakes aren’t cute demos; they’re liability.

V7 Go, per OpenAI’s write-up, uses GPT‑5.6 Luna to extract from large file corpora into a Context Graph of entities, relationships, and cited evidence, then GPT‑5.6 Terra and Sol for heavier reasoning and multi-step workflows. The pitch is institutional memory agents can query instead of re-searching the haystack every time.

We’re inclined to like the problem statement. Long-context alone is a blunt instrument: expensive, slow, and still bad at “which version is live.” A graph that connects facts to source documents, updates when new files land, and exposes MCP search into ChatGPT and Codex is the architecture enterprises keep asking for. Founders Rizzoli and Edwardsson started V7 in 2018 after a computer-vision accessibility app; V7 Go is the agentic workflow layer on top of that bet.

The performance claims are company-measured, so we treat them as signals, not gospel. V7 says agents complete 50–100 step workflows in minutes at 99.9% accuracy with an auditable trail. On HERB, a benchmark for connecting information across enterprise systems, V7’s retrieval-only system beat the official baseline by 69% and cut hallucinations on unanswerable queries by 38%. Customer-flavored numbers include asset managers screening deals 21× faster (a full day to about 15 minutes), a financial services team cutting review from more than 100 hours to under 10 (about $12,000 in expert cost per task), and insurance teams cutting claims-processing errors by 13.5% versus a manual baseline after agents got historical claims and policy knowledge.

On the model side, V7 maps steps to fast/medium/smart tiers: Luna for high-volume extraction, Terra or Sol for chat and tool-heavy paths. It reports GPT‑5.6 Sol dropping tool-call error rate from 2.7% (GPT‑5.5) to 0.2% on its Context Graph benchmark, GPT‑5.6 Luna at 78% lower cost per document than GPT‑5.4 mini, and some multi-call workflows finishing up to 50% faster. When Sol saturated older tests, V7 built harder graph-query questions; GPT‑6 Astra hit 89% on the hardest set in that harness.

Advocate the useful part: source-linked memory plus tiered models is how agents become coworkers instead of amnesiac interns. Name the downside: if the graph is wrong, every downstream step inherits the error — faster, with a nice citation to the wrong PDF. “99.9%” and “mission critical” in the same breath demand independent audits, especially in underwriting and deal rooms. Access to SharePoint and Drive also means a compromised agent is a compromised records system.

I’m watching V7’s next step: workflows that fire when graph facts change and flag stale analyses. That’s productive automation. The trajectory if this becomes normal is enterprises competing on context quality, not agent count — Rizzoli’s own line — which is healthier than a swarm of forgetful bots. Just keep humans on the exceptions.

OpenAI’s customer story “How V7 gives AI agents institutional memory” describes V7 Go’s Context Graph, GPT‑5.6 Luna/Terra/Sol usage, MCP exposure to ChatGPT and Codex, and GPT‑6 Astra graph-query tests. Figures above are as stated by V7/OpenAI in that piece.