An agent that gets smarter where it counts: your device, your regulatory journey, and all of our experience.

Most compliance software starts cold. Every session you re-explain your device. You restate the pathway you picked, the market you're going after, the thing you decided three weeks ago. The tool  forgets, so you have to carry the context. A generic AI tool has a different problem. At best it can read the files you upload or remember fragments from a chat. That still only gets you halfway there. It doesn’t know why your team chose a pathway, what an auditor pushed back on, or what your regulatory lead decided in the meeting where the real work happened.

Dovetail's agent doesn’t work that way. It remembers and draws on everything that matters. Your meetings, your documents, and the decisions behind your submission, plus everything our regulatory experts know from years of experience.

How Dovetail’s memory works 

At a high level it's simple. The agent captures what matters from the work you're already doing and brings the relevant pieces back when you need them.

Two sources feed all of it: our regulatory experts and you. Nothing here is a model we retrained on your files overnight. It's an agent that holds context and gets to use it, the way a good teammate would after a few months on your account.

That's the whole idea. The interesting part is what it makes possible.

Good on day one: it already knows the craft

When you start with Dovetail, your agent is not starting from zero. It already carries how our experts think. Our experts feed it continuously with learnings from our day-to-day operations. So the agent improves every day for everybody.

Our regulatory experts make hard calls constantly. How to classify a borderline device. How to structure an evidence argument that survives scrutiny. What a notified body is going to push back on before they push back on it. The agent picks this up from the calls they're on and the feedback they give in writing, as the work happens, not from someone stopping to document it afterward. That kind of knowledge is craft, and the agent carries it.

The practical effect: you get the accumulated judgment of a regulatory team that does this for a living, on day one, without paying for it to be relearned on your time. This is the expertise that feeds your agent, and it's the thing a bigger model can't give you on its own. Before any of it reaches the agent, a human reviews it, so what gets through is real expert knowledge, not noise.

Sharper over time

Day one is the expert layer. Everything after that is yours. The more you work with the agent, the more it knows about your specific device, and that shows up in three ways you will feel immediately.

You stop re-explaining your device. The agent remembers your strategic decisions, what you tried during development and testing, and what worked and what did not. It remembers why you decided against running a trial, or why you chose a specific predicate.

It stays consistent across a long time horizon. Regulatory work runs for months and passes through a lot of hands. The agent remembers the pathway you chose, your predicate, your target markets, your timeline. Ask it something in month six and the answer lines up with the decision you made in month one. Whoever on your team is asking, the agent gives them the same grounded answer, so your program doesn't drift every time a new person picks it up.

It learns how your team works. Your terminology. The voice you write in. How you like documents structured. The agent adapts to how your company operates instead of making you adapt to it. A small thing per interaction, a large thing over a full program.

This isn't the memory you've seen this year

You've watched ChatGPT and Claude ship memory. That memory is useful. It remembers you as a user: your role, your preferences, the way you like answers. But it does not carry your regulatory program. It does not know the meeting where your team rejected one predicate and chose another. It does not know the objection your auditor raised, the evidence gap your expert found, or the decision that changed your submission strategy.

That is the risk with generic AI in regulated work. It can sound confident while missing the thing that matters. It can give a clean answer built on incomplete memory. In QA/RA, that is not a productivity issue. It is how programs drift, documents contradict each other, and teams walk into an audit with a story they cannot defend.

Dovetail's memory learns from the work itself, not from a chat window. The decisions made on a call with our experts. The feedback they give in writing. The facts your team settles along the way. It holds your regulatory program, your device, your evidence, your timeline, and it is fed by people who do regulatory day-in-day-out. It also lives under privacy a consumer chatbot does not offer, because the data going into it is the kind you have to account for. Same idea, different job. The job is the part that matters.

And it stays yours

Here's the part a regulated founder should care about most. Everything the agent learns about your program is scoped to either you or your company. It never becomes shared expert knowledge. It never reaches another company's agent. And it never trains a model, ours or anyone's.

There's no black box about what it's using, either. Every memory the agent pulls carries its origin: where it came from and when. So when it leans on something it remembers, a decision you made or a rule from our experts, that line can be traced back to its source. You're not left guessing why it said what it said. That's the difference between a tool you can audit and one you have to trust blindly.

It's also where the regulation is heading. The EU AI Act's data-governance rules for high-risk systems are coming into force, and they ask exactly the questions you should be asking any partner you hand regulated data to: where did this data come from, who can see it, can you prove it was handled properly, can you remove it. We can answer all four, because the memory was built as records from the start, not as an afterthought bolted onto a black box.

The human layer

The agent is useful because it has memory. It is more useful because it sits inside an expert-led system. Our regulatory experts still make the hard calls. The perfect coming together of machine intelligence and human judgment. Envoy gives the QMS layer a proper home. The agent keeps the context moving between them, so the plan you agreed on in a call can show up in the SOP, the evidence review, the audit trail, and the next question your team asks. AI can help you prepare for an auditor. It cannot sit in the meeting for you. That is why we built this around experts, records, and a QMS, not just a chat window.

What this means

The point is not memory for its own sake. The point is a regulatory partner that gets more useful every day. 

You spend less time re-explaining. Your team gets more consistent answers. Your documents line up with the decisions behind them. Your experts have better context. And your audit trail is clearer because the system remembers where the knowledge came from.

Severin Högl is the CTO and co-founder of Dovetail (formerly FormlyAI), and has experienced the difficulties of medical device CE marking firsthand, building and certifying a patient monitoring app in a previous venture. Today he leads the team behind Dovetail's AI agent, building regulatory AI that's a step ahead of the generic tools everyone else is using.