Your agents are starting to remember. Until recently, an AI agent treated every conversation as its first. You could tell it the same thing ten times and it would learn nothing. That is now changing. The major platforms are shipping memory as a built-in capability. Databricks released managed memory for agents in September, and followed in October with pipelines that review an agent’s conversations, distill the facts worth keeping, and save them so the agent can use them next time. 

For the business, this is mostly good news. An agent that remembers is simply more useful. A patient services agent that recalls last month’s case does not make the caller start over. A medical inquiry that stretches across weeks of follow ups can finally be handled as one continuous thread instead of five cold starts. 

But notice what just happened. Your organization now has a system that creates data on its own. Each remembered fact is a new record that no employee typed in, no form validated, and no one reviewed. The AI decided what was worth keeping, wrote it down in its own words, and stored it. Multiply that across every agent and every conversation, and you have a fast growing collection of information that nobody created.

That matters especially in healthcare and life sciences, because of what an agent might reasonably decide to keep. Consider two memories that look harmless on the day they are made. A caller misspeaks, or the agent mishears, and a wrong detail is distilled into a stored fact. Every later conversation quietly builds on it. Or a patient mentions her husband’s condition while scheduling her own appointment. The agent now holds health information tied to the caller. Neither memory was created with bad intent. Both are now records that no one reviewed, and no one owns.

The downstream questions are familiar ones. Who is allowed to see what an agent remembers? How long is it kept? If a patient asks for their information to be corrected or deleted, does that request reach the agent’s memory too? If one agent shares what it knows with another, do the access rules travel along? And when an auditor asks where a piece of information came from, can you answer?

Here is the reassuring part. None of this requires inventing a new discipline. Your organization already knows how to govern data: classify it, control who can access it, set retention, track where it came from. Agent memory is simply a new place that discipline needs to reach. In Databricks, memory lives inside the same governed data platform as everything else, so the controls you already use can extend to it. We believe the companies that treat memory as governed data from day one will be the ones comfortable letting their agents take on more.

The risk is not the technology. The risk is treating memory as a technical detail each agent team sets up on its own. Do that, and a year from now you may discover an unmanaged store of your most sensitive conversations, built quietly and in good faith, one remembered fact at a time. 

These capabilities are new and still in Beta, so the right posture is plan, not deploy. Decide on purpose what your agents should remember, and write it down before the system decides by default. Treat remembered information as seriously as the systems it was learned from. Put a retention clock on it before the first fact is written. 

An agent that forgets everything is a worse product. An agent that remembers everything is a liability. The difference between the two is not a technology choice. It is governance.