We made our AI forget everything every 30 days. Users started sharing more.
When we first built Murror, we stored every journal entry forever. More data meant better emotional pattern detection, smarter insights, richer reflections.
Our AI kept getting better. Our users kept getting quieter.
We dug into the data. Users who'd been on the app for 3+ months were writing shorter entries. Less personal. More surface-level. "Had a good day." "Feeling okay."
We ran user interviews. The pattern was clear: the longer people used Murror, the more aware they became that they were building a permanent record of their most vulnerable moments. One user said, "I want to be honest, but I also know this is all being saved somewhere."
So we tried something radical: a 30-day memory window. After 30 days, raw journal entries get summarized into anonymous emotional patterns, then the original text is deleted. Users can pin specific entries to keep them — but the default is forgetting.
The team thought I was crazy. "You're throwing away our best training data."
What happened:
Average entry length increased 3x
- Users mentioned specific people by name 5x more often
- 30-day retention went from 23% to 41%
The lesson: in emotional technology, the right to be forgotten isn't a privacy feature. It's a trust feature. People will only be truly honest with technology that promises not to remember everything.
Has anyone else experienced this tension between data retention and user trust?
Replies
the pin feature is the part I'd want to understand better - if a user can pin an entry to keep it forever, doesn't that just recreate the same "this is a permanent record" anxiety for whatever they choose to pin? seems like the honesty boost you're seeing comes specifically from knowing the default is deletion, so I'd guess people pin way less than they'd expect to going in. curious if that held true, or if pinning became its own habit once people got comfortable with the 30 day window
This is super cool. I have a digital health app and think this could be beneficial. It also helps with shortening the prompt fed back into the model!