Your AI Agents Now Remember Your Data. Who Decides What They Forget?

Most organizations believe the hard part of governing AI agents is controlling what they can open: which apps, which folders, which permissions.

The larger problem is what agents keep. At Gemini at Work last week, Google introduced coworker agents with their own email, calendar, Drive and "persistent storage," plus a semantic memory Google describes as "a structured knowledge base it builds as it reads documents, talks to people, and works with other agents." OpenAI's always-on Dots, launched late last month, also gradually learn an employee's preferences, standards and working habits.

Every one of those memories is a new copy of your business knowledge: summarized, reinterpreted and stored somewhere your catalog has never seen. Nobody has given it an owner, a retention rule or a lineage trail.

And it multiplies. Google says SOMPO alone runs over 10,000 custom agents across 34,000 employees. Every agent that remembers is another knowledge base growing quietly beside the ones you already govern.

Why Is Agent Memory a Data Problem, Not an AI Problem?

For years, your data estate had a recognizable shape: systems of record, a warehouse, shared drives, and the copies people made of them.

Agent memory is different. It is not a file someone exported. It is a set of conclusions an agent drew from your data, stored in a form your teams never designed.

Picture an agent that read last quarter's pricing deck in its first week. In week six, pricing changes and the deck is updated. The agent's memory of the old discount rules is not, unless someone knows it is there.

That is the stale-copy problem data teams have fought for decades. It is now happening inside systems that act on what they remember.

The Policy Is Only as Good as the Labels Under It

Google says every agent gets its own cryptographically attested identity, and that you can write a policy once, such as "agents may not open documents classified Need to Know."

Read that example closely. It only works if the document is already classified.

Google's own keynote notes that ninety percent of enterprise data is unstructured. In most organizations, much of it has never been labeled, owned or reviewed. A policy that blocks sensitive documents cannot block what nobody marked as sensitive.

None of these questions are particularly difficult. The problem is that most organizations answer them by hand, once, during an audit, while agents learn continuously.

"It Was the Agent" Will Not Satisfy Your Auditors

In remarks reported today, Entrust COO Anudeep Parhar argued that "it wasn't me, it was an agent who did it" will not satisfy auditors or regulators. Consequential actions, he said, need evidence of what happened and why it was authorized.

Action logs help. They record what the agent did. They rarely record what the agent knew when it did it, or where that knowledge came from.

And visibility is already thin. IBM research cited by Dataiku found that fewer than one in five organizations maintain a complete, current inventory of their AI systems. If you cannot list your agents, you cannot list what they remember.

How to Get Your Data Ready for Agents That Remember in 30 Days?

You do not need to pause your agent rollout to fix this.

Within the first month, focus on the sources your first coworker agents will actually be given: shared drives, team sites, and the handful of tables behind your most-requested reports.

Start by discovering and classifying what sits in those sources, so policies like "Need to Know" have labels to act on. Then assign a verified owner to every source and flag what has been superseded.

Next, map lineage from each source to the reports, workflows and agents that depend on it, so that when a source changes or is retired, you know which agents need to be refreshed.

DataManagement.AI accelerates this process by automatically discovering metadata, mapping lineage across the modern data stack, identifying downstream dependencies, and creating a centralized knowledge layer that stays continuously updated as your environment evolves. Ownership and policy travel with every entity, so the rules you set follow your data wherever an agent picks it up.

Instead of spending months labeling files by hand after an agent has already read them, your teams can give every agent a classified, owned and traceable starting point, and know exactly what to correct when a source changes.

Agent Memory Is Actually a Data Lifecycle Problem

Memory is what makes these agents useful. An agent that remembers your standards does better work on day sixty than on day one.

But memory turns every agent into a data store, and every data store needs the same things: an owner, a source of truth, a retention rule and a way to be corrected.

Organizations that scale AI agents safely do not just control what agents can open. They govern what agents learn, and what happens to it when the truth changes.

Warm regards,

Shen and Team