The complaint is everywhere in AI companion communities: "I told it something three weeks ago and it forgot." So when solo founder Jake (who publishes under his first name) built Dusk AI, he made persistent long-term memory the core feature. Every conversation would build on the last. The AI would remember what you shared about your life, your struggles, your name for your cat.
Then he watched what users actually did with it.
The Gap Between Stated Preference and Behavior
Users said they wanted memory. They praised it in feedback sessions. But engagement data told a different story. People who had extensive memory profiles - dozens of stored facts about themselves - weren't necessarily the most active users. Some of the most engaged users had light memory footprints.
This pattern shows up in product design research outside of AI too: what people say they want in a survey and what they return to daily are often different things. In companion apps specifically, the appeal of memory might be about the feeling of being known rather than the functional use of specific recalled facts. An AI that responds warmly and attentively in the moment can create the same sensation as one reciting accurate biographical details.
There's also a darker possibility the founder raised directly: some users may not actually want continuity. A companion that starts each conversation fresh offers a kind of social reset that a human relationship doesn't. No accumulated disappointments, no context carrying over from a bad day three weeks ago.
What This Means for AI Product Design
For developers building AI companions, this is a useful data point against feature-first thinking. Long-term memory is technically hard to implement well - it requires deciding what to store, how to surface it naturally, and how to handle outdated information without being creepy or wrong. If that engineering investment isn't actually driving retention, it may be the wrong priority.
The more defensible product insight from Dusk AI's experience is probably this: presence matters more than recall. An AI companion that feels engaged and responsive in the current conversation likely outperforms one that accurately remembers your sister's name but responds with generic warmth.
That said, Dusk AI is one app with one founder's read on the data. Sample sizes in early consumer AI apps are small, user populations skew toward early adopters, and "engagement" can mean many things. The conclusion that memory doesn't matter would be premature. The conclusion that memory alone isn't sufficient - that one seems solid.