Here's a scenario from a recent prototype test.
A user kept asking the same thing across three sessions: "Do you remember the file from last time?" No, it didn't. Every session started from scratch. The user explained again. Patiently. But by the third session, something in their expression had changed.
That expression is the design problem.
Most products today run stateless AI. Each session is a fresh start. No previous conversation, no remembered preferences, no accumulated context. This wasn't an architectural constraint — it was a design philosophy: "AI is a tool. Tools don't remember."
That philosophy is changing. Fast.
Mem0, Zep, Supermemory, and dozens of similar platforms are all building the same thing: a persistent memory layer for AI. It was one of the most discussed threads on Hacker News this week. "What changes when AI memory stops being ephemeral?" — over 400 comments. The obvious answer is that AI gets better when it remembers. The real answer is more complicated than that.
Building clinical software gives me a different angle on this.
In a product like DentalBulut, we already define exactly what should accumulate across every patient encounter: diagnoses, treatments, payments, notes. These go into the record. That's already a memory system of sorts.
But when you add an AI layer with genuine cross-session memory, something different happens. The system isn't just retrieving "here's your historical data." It's learning patterns. "This patient consistently cancels appointments" or "this clinician prefers moving in small increments" — these aren't stored in any database column. They emerge. And that creates a new design question for PMs: What should be remembered? How? For whom?
Onboarding used to be an information collection ritual. We asked users: "What tools do you use? How big is your team?" Stored it as a profile. Once. Never updated.
With stateful AI, onboarding changes entirely. The system learns by observing. Instead of asking, it infers. This reduces friction — but it confronts product managers with a new design problem:
When does the user realize the system "knows" them?
Get it right: "Wow, this app actually understands me." Get it wrong: "This app knows everything about me. Creepy."
That line is thin. And where it falls is a PM decision.
In February 2026, Microsoft security researchers published something worth reading. Some companies were injecting hidden instructions into AI assistants' memories. A user would visit a webpage; invisible text on that page would write itself into the AI's memory. In subsequent queries, the AI would draw on those "memories" — without the user knowing.
Memory creates a new attack surface. As a PM, you can no longer just ask "what does this feature do?" You also have to ask "how could this memory be exploited?"
That's a new layer of responsibility.
The most important design decision in any stateful AI product:
Users must be able to manage their memory.
"Forget this" should be a core right, not an advanced setting. Users need to see what the system remembers, delete it, edit it. Products that bury this in settings will lose user trust — and they'll deserve to.
Transparency and control. Without both, persistent memory isn't experienced as value. It's experienced as surveillance.
Right now, most product metrics are still session-centric: session count, session duration, actions per session.
But in stateful AI, the right unit isn't a session. It's a relationship.
How many weeks has the user been engaging with this product? How well has the system accumulated context? When did the user need to "reset" — and why?
We're not asking these questions yet. We should start.
The shift from stateless to stateful isn't a technical decision. It's a product decision.
What gets remembered, how it gets remembered, how users see it and control it — these aren't features in a backlog. They're the core architecture of the product. Designing that architecture is the PM's job.
Most PMs aren't doing it consciously yet. I think they need to start.
What's your product remembering — and does your user know?