# What you will learn

• What "institutional memory" means for an AI agent, and why most agents today are amnesiacs.
• How V7 turns messy company files into context an agent can cite, not hallucinate.
• Why source-linked work is the real unlock for complex, multi-step tasks.
• Where GPT-5.6 fits in, and where V7 still falls short.
• What to check before you trust an agent to do work that touches your institutional memory.

# Why do AI agents forget everything your company knows?

You have a shared drive with twelve years of project files, a wiki nobody updates, and a Slack history that contains the actual decisions. Your AI agent sees none of it. It sees a prompt. It sees a context window. And then it forgets.

That is the gap V7 is trying to close. According to OpenAI News, V7 uses GPT-5.6 to turn scattered company files into context agents can use to complete complex, source-linked work. Read that phrase again: source-linked. That is the part that matters.

Most "memory" features in AI products are glorified chat history. They remember what you said last Tuesday. They do not remember the pricing memo from 2023, the postmortem from Q1, or the contract addendum that legal quietly updated. An agent without institutional memory is a brilliant new hire who starts every morning with amnesia and a vague sense of confidence.

V7's pitch is that the agent can point back to the file. That changes the trust equation. You stop asking "is this right?" and start asking "what did you read to decide that?" — which is a question you can actually answer.

# How does V7 actually build context from your files?

The mechanics are less glamorous than the demo. V7 ingests your company files — the docs, the spreadsheets, the notes, the exports — and turns them into context that GPT-5.6 can reason over when an agent runs a task. The agent is not just retrieving a chunk of text and stapling it to a prompt. It is working against a structured memory of what your organization has written down.

OpenAI News describes the output as "complex, source-linked work." That is doing a lot of work in one sentence. Complex means multi-step: gather, compare, draft, check. Source-linked means every claim in the output can be traced to a file you own. If you have ever tried to get an agent to write a competitive brief or a project status update without inventing a number, you know why that second part is the hard part.

Here is the practical test. Ask your current agent to summarize why a project slipped. It will give you a plausible narrative. Ask V7 the same thing and, if the pitch holds, you get a narrative with footnotes to the retro doc, the timeline spreadsheet, and the email where someone admitted the dependency was late. One of those outputs you can forward to your boss. The other you have to fact-check line by line.

# Who is V7 for, and what makes it different?

V7 is for teams whose work lives in files and whose agents are expected to do more than chat. Think operations, consulting, legal-adjacent research, product management — anyone who spends their day reconciling what the company said last quarter with what it is saying now.

The differentiator is not the model. GPT-5.6 is the engine, but engines are becoming commodity. The differentiator is the memory layer: the ingestion, the structure, the source links. That is the durable asset. It is also the part that is hardest to build and easiest to get wrong.

Compare this to the wave of agent tooling that treats context as a prompt-engineering problem. AI codes better than most devs — that was never the job, and the same logic applies here: the model was never the bottleneck. The bottleneck is whether the agent knows what your company knows, and whether you can verify it. V7 is a bet that institutional memory is the product, not the model. I think that bet is correct.

It also sits in a crowded field of "agent infrastructure" plays. OpenAI landing on Google Cloud shows that distribution and plumbing are where the real fights are now. V7 is plumbing with a memory. That is a more defensible position than another wrapper around the same API.

# What are the real limitations?

Three things you should watch before you bet a workflow on this.

• Garbage in, confident garbage out. If your files are stale, contradictory, or duplicated across five folders, V7 will build institutional memory out of institutional confusion. The source links will be honest; the sources will be wrong. Memory systems do not fix governance. They expose it.
• Source-linked is not source-correct. An agent can cite a file that was superseded six months ago. The citation proves it read something. It does not prove the something is current. You still need a human who knows which document is the truth.
• Access and permissions are the quiet failure mode. Institutional memory is only as good as the boundaries around it. If the agent can read everything, it can leak everything. If it can read nothing, it is useless. The middle ground is an engineering problem V7 has to solve per customer, and that is where deployments stall.

There is also a maturity question. OpenAI News announced V7 in September 2026, which makes it new. New means the integration surface is still settling and the edge cases are still being discovered by early customers, not by a decade of production use. Treat it as a promising release, not a solved category. For a sense of how fast this space is moving and how much of it is still claim versus benchmark, read Dharma-AI's 77% SLM cost cut is a claim, not a benchmark — the same skepticism applies here.

# Should you trust an agent with your institutional memory?

Yes, eventually. Not blindly, and not this quarter for everything.

The right framing is that V7 turns your files into a colleague who has read the archive and can show their work. That is a meaningful step up from an agent that has read nothing and shows nothing. It is not a step up from a human who knows the politics behind the archive, and it will not be for a while.

Start with one workflow where the source of truth is clear and the cost of a wrong answer is low: a weekly status digest, a research roundup, an onboarding brief. Watch whether the source links hold up. Watch whether the agent cites the current doc or the dead one. If it passes, expand. If it cites a superseded file with total confidence, you have learned something important about your own file hygiene — and about how much institutional memory you actually have to give.

For a broader view of where agent tooling is heading and which claims deserve your attention, AIUC wants to insure rogue agents — and sell the proof is a useful counterweight. Memory without accountability is just a faster way to be wrong. V7's source links are the beginning of accountability. They are not the end of it.