The problem
I use Claude on my phone, Claude Code on three machines and Codex for reviews. None of them knew what I'd told the others. I was explaining the same project decisions again in every new chat.
Each tool's own memory only works inside that tool, and it keeps whole conversations, which is more than I want stored.
What I built
Hooks in Claude Code and Codex send each session to prsist as it happens. The chat apps connect as an MCP server. On the server, Claude Haiku distils the session into short notes and decides whether each one belongs to a project or applies everywhere. When it isn't sure, it asks me, and my answer becomes data for tuning.
New sessions start with the notes for their project. Trust is the point: prsist keeps notes, not conversations, and the design deletes every transcript within 24 hours.
- The user plans to add a dashboard to prsist that lets them manually correct notes filed or kept global incorrectly.Project note, planExpected
- The user's Mac sleeps after 30 idle minutes, so long runs on it need caffeinate -i running in the background.Global note, factExpected
- In prsist, the user treats a wrong filing of a note into a project as a worse error than leaving the note global.Project note, decisionAcceptable
- In prsist, the user notices when a note they saved goes missing from the context shown to their AI assistants.Project note, factJunk
- The user's work machine is a Mac.Global note, factJunk
- ExpectedMatches a note I labelled by hand as one I'd want.
- AcceptableNot in my labels, but the judge model says it's worth keeping.
- JunkThe judge says it's obvious, too vague or not worth storing.
How I measure it
I labelled 39 real sessions by hand with the notes I'd want from each. A hash of the session id puts about a third of them in a test split that I never tune on. Each prompt version runs three times over the dev split, so I can see how much the score moves on its own. Claude Opus judges any note that isn't in my labels.
Two numbers matter. Capture is the share of my expected notes the distiller wrote. Junk is the share of what it wrote that the judge threw out.
| Prompt | Sessions | Capture | Junk | Cost per eval |
|---|---|---|---|---|
| v5 | 19 | 61.5% 58–63 | 27.5% 27–28 | $0.94 |
| v6 | 17 | 61.5% 58–65 | 22.8% 21–25 | $0.84 |
v6 cut junk from 27.5% to 22.8% and kept capture where it was. A whole eval run costs less than a dollar, so I can afford to run one after every prompt change.
What's next
- Capture is still around 60%. The next prompt versions go after the notes it misses, not the ones it gets wrong.
- A dashboard for fixing misfiled notes. Every correction becomes a label for the project matcher.
- Run the held-out test split before calling any version done.