Personal AI memory is not company memory

Somewhere in your firm, someone's ChatGPT knows a lot about your business. It knows your main clients, the way you price, the tone you like in emails. It learned all of it from one person, over months of chats.

When that person leaves, it leaves with them. When they switch to Claude because a colleague said it writes better, it stays behind in ChatGPT. When the memory fills up, it starts dropping things, and on OpenAI's own community forum users have reported exactly that happening with no warning. In November 2025 some reported their saved memories wiped entirely.

That is personal AI memory working as designed. It is built to remember a user. A company needs something else.

What the vendors now offer

Over the past eighteen months every major AI provider has shipped some form of memory.

• ChatGPT added recall from past chats in April 2025, on top of saved memories, and later added company knowledge connectors that search tools like Slack, SharePoint and Google Drive.
• Claude added memory for individuals and then, in September 2025, project-scoped memory for Team and Enterprise plans.
• Microsoft has Copilot Memory, and announced Work IQ at Ignite in November 2025, built over Microsoft 365 data.
• GitHub put agentic memory for Copilot coding agents into public preview in January 2026, scoped to a repository.

These are real features and they solve real annoyances. Re-explaining yourself every session is tedious, and vendors heard that. But read the list for scope. One user. One project. One repository. One vendor's ecosystem. Search across documents. None of it is designed to hold what a company has decided, apply it to every person and every agent equally, and let the company see and change what is held.

Four gaps, and why they matter at work

It belongs to a person, not the firm. A company's rules should not depend on which employee happened to teach their assistant. Ten staff with ten personal memories means ten slightly different versions of your prices, terms and voice.

It belongs to one vendor. Memory built inside ChatGPT stays in ChatGPT. Most firms use more than one AI tool, often without meaning to. The moment you change tools, or a tool changes its terms, the memory does not come with you. We learned this one directly: a stretch without access to one AI vendor meant our setup had to work on others.

You cannot fully see or audit it. Users report memories they never asked for and memories that disappeared. A company needs to know exactly what its AI is working from, who changed it and when, and to be able to correct it.

Nothing enforces it during the work. Instruction files and custom instructions exist, but in long sessions and long agent runs they fade, a problem we have written about several times since September. Memory that is loaded once and then trusted to persist is not much protection.

MIT's 2025 report on business AI, better known for its headline failure rate, made the point we think matters most: many tools fall short because they do not retain feedback, adapt to context or improve over time. Personal memory remembers preferences. It does not hold the corrections a firm has paid for.

What we changed today

At Provecta Group we have run our own work on a written company memory since the summer: rules, verified facts, voice, approvals and lessons, each correction written down with its reason and how to apply it. Today we finished moving all of it into a single store the company owns, and took the last of it out of any one vendor's settings.

How it works now:

• One root. Every AI session and every scheduled agent starts at one root file and loads the rest in a fixed order: the root instructions, then the company rules, then the memory index and the memory entries it points to, then the role profiles, then the runbooks for the task.
• Any model. Each AI tool's own settings file only points back to that root. Tools that can read the store directly do so. Web tools that cannot get a redacted copy of the same files. Changing models does not mean starting over.
• Written corrections. As of today the memory holds 195 entries besides its index, and 42 of them are rules that came from a correction. About half of all entries carry both a written reason and how to apply it, including 40 of those 42. There are 26 role profiles, each paired with an advisor view of what someone experienced in that role would check.
• History. Every change is versioned, so we can see when a rule appeared and why.

A small example of the loop. Two days ago, an agent working on a batch of outreach notes rewrote notes that had nothing wrong with them, wrote one that referred to a conversation that had not happened, and produced a batch that all followed one template. The correction became a root rule the same day: check the real state before changing anything, fix only real faults, never state an event that has not happened, and read a batch side by side for sameness. Every session since has loaded it.

What it does not fix

We would be misleading you if we stopped there.

A written memory does not make a model obey. It makes drift less likely. Anything that sends, spends, deploys or deletes still needs a gate the AI cannot talk past. In our own setup, a short list of routine actions can go out on their own, each with a written approval, a date and a reason, and everything else waits for a yes.

It does not stop the model inventing things outside what the memory covers. It narrows where that can happen.

It makes security harder, not easier, unless it is governed. A memory every agent reads is a concentrated copy of what the firm knows. IBM's 2025 breach report found shadow AI in 20% of the breaches it studied. Ours is scoped by company, keeps credentials out entirely, and is redacted before any copy leaves for a web tool. That is ongoing work, and we check it ourselves rather than through an outside audit.

It is labour. Someone has to decide what is true, write it down and keep it current. A stale memory gives confident wrong answers, repeated across every session.

And we have not measured it as a controlled before and after. We run our company on it because the failures it targets are well documented, not because we can quote a percentage.

Where to start

Ask three people in your firm to show you the custom instructions or saved memories in their AI tools. Read them side by side. Where they disagree on a price, a term or a rule, you have found the first entries for a company memory, and the first sign of what personal memory has been quietly doing in your firm without you.

Book 20 minutes at pgco.world/book.