Selected work / Case study 02
Memory you
can trace.
Saving one useful idea, finding its original words later, and keeping the conversation easy to use.
- My role
- Product direction, interaction design, requirements, acceptance criteria
- Part of
- StephOS Archives, the foundation for Continuity with Consent
- Built with
- AI coding agents directed by me
- Status
- Prototype I use myself; no external users yet
A useful idea shouldn't
lose its original words.
In long AI conversations, the best ideas surface in passing. Keeping one usually means copying it somewhere, or asking the assistant to "remember" it and trusting a paraphrase later. I wanted the person to keep the exact words, with their source, without leaving the conversation.
The person
Someone doing long-running AI-assisted work who wants to keep an idea without derailing the conversation.
The task
Save a passage, return in a different chat, get the exact wording back, and see where it came from.
The challenge
Make evidence available without turning every conversation into an admin panel.
Three things AI
products blur.
Most of the design came down to keeping three ideas apart, in the data and in the interface.
Saving is not approval.
Keeping an idea doesn't make it approved memory. It stays a saved idea until the person decides otherwise.
Approval is not truth.
A person can approve a record. That doesn't make every statement in it true.
Quotation is not interpretation.
Original words come from what was saved; the model's explanation stays separate. A model saying an action ran is not evidence that it did.
Less in the way.
The first version put "Keep this idea for later" under every message. It worked, and it made every conversation feel like a form. I moved it into collapsed message details, one tap away, and added a review step so nothing is saved by accident.
Before: on every message
Keep this idea for later
After: one tap away
Review before saving
Illustrative wireframes with fictional content
A finding that changed the design
An assistant once told me a saved idea needed approval to last. It doesn't. The model was explaining a status it didn't own, so now the app, not the model, states what's saved and what's approved.
Find the words.
Let the person choose.
From your saved sources
Original wording · Saved idea · Not approved memory
- 1
One match
Show the original words, exactly as saved, with their source and status. No model is asked to repeat them.
- 2
Several matches
Ask the person to choose, then recheck the chosen source before showing it.
- 3
No valid match
Say so plainly. Never substitute generated wording for the original.
Illustrative with fictional content
Make "done" observable.
Every requirement came with a check I could run myself. "Done" isn't a feeling.
| Requirement | How it's checked |
|---|---|
| Exact save | The saved text matches the chosen passage, character for character. |
| Cancel means no save | No save request is sent and no new record exists. |
| Approval stays separate | Approved memory and current work are unchanged by saving. |
| Literal recall | In a different chat, the same words come back, with no model call. |
| Safe ambiguity | The person chooses between matches; stale or changed ones are refused. |
| Honest failure | The limit is reported. A quotation is never invented. |
From one saved idea to
memory with consent.
Archives proved the foundation: original words, kept and retrievable on purpose. The next question was automatic recall that the person controls, which became Continuity with Consent.
Proven in my own use
- Exact passages saved after review
- Original wording returned in a different chat
- Saved, approved and quoted kept distinct
Not yet
- No external users yet
- Screens shown here are illustrative
- Imported history is out of scope
Next case study
How I ship
with AI agents.
A designer taking ideas to installed, working software, with one AI building and another checking.