A private iOS journal whose AI cites only your own entries
AI journaling apps answer with generic advice that can drown out the writer’s own voice. I wanted reflections built only from what the writer had written.
I designed and built Memento, a native iOS journal whose local retrieval pipeline cites past entries in every reflection.
Early testers wrote more often once they could see where each reflection came from.
People stop journaling when the AI feels like it is watching them.
Generic model responses flatten the writer’s voice, and insights with no sources feel intrusive. Either one is enough to make people stop writing.
Three limits shaped the product: first, private by default; second, answers only from the user’s entries; and third, I was the only designer and engineer.
A journal is only useful if people keep writing in it.
Reflections depend on a long history of honest entries. If writers do not feel safe, they stop, and the AI has nothing to work with.
I limited what the AI could say, and showed where each line came from.
Retrieval runs only over the user’s journal
Memento finds related entries first, and the model answers only from those. Reflections reuse the writer’s own phrasing instead of generic advice.
Instead of A general model that answers anything, with the journal passed in as extra context.
Trade-off With only a few entries, Memento has little to reflect on.
A Dive Deeper panel lists the source entries
Each reflection links to the exact entries it used, so the writer can check it against what they wrote.
Instead of A summary with no visible sources.
Trade-off An extra panel on a screen I wanted to keep almost empty.
Left: the entry feed, with no scores or streaks. Right: Dive Deeper, where reflections link back to the entries they used. No streaks and no push notifications
After each entry, the AI asks one follow-up question. Tapping it opens a blank page for the next entry.
Instead of Streaks and daily reminders, the standard retention tools for journaling apps.
Trade-off Without them, people come back only if writing feels worth it.
The AI stays off screen until you save, and asks when history is thin.
- WritingNo AI on screen until the entry is saved.
- ReflectingFinds related entries, then asks one question.
- Grounded answerShows the reflection with links to its source entries.
- Not enough historyAsks an open question instead of making up an insight.
Testers wrote more once they trusted that entries stayed on the phone.
In early testing, people wrote more often and at greater length once they understood their entries stayed on the device. Memento is in closed beta ahead of public launch.
Next: make the first week useful with only a few entries.
Reflections improve as history grows, so the first week is the weakest. I am designing prompts that help new writers build up enough entries for reflections to be useful.