You, several machines, several projects, and more than one AI instance. Claude Code on the desktop, another on the laptop, a chat session on your phone, maybe Codex when you're feeling adventurous.
Here's the shape of the problem. Monday's desktop session spends forty minutes working out why the migration deadlocks. Wednesday's laptop session hits the same wall, knows nothing about Monday, and spends the forty minutes again. Multiply that by every project you've touched this quarter.
Nothing you know is durable, because your instances have no shared place to put it.
Set up once
Sign in, then connect each machine's assistant with the one-paste setup - the whole thing is in Getting started. It's keyed to your account, so every instance you connect sees the same workspace. No syncing, no files to copy between machines.
Use one project per repo. It's the setting that keeps recall precise, and projects are free.
Monday: the instance learns something, and keeps it
You're deep in a deployment problem. Your instance finally works out what's happening. Rather than letting that die with the session:
You: That's the answer. Remember it for the
acmeproject - the migration has to run before the worker boots or it deadlocks on the advisory lock. Type it as a gotcha.
Claude: Stored as "Deploy gotcha" on
acme, typedgotcha.
Two seconds of work. That's the whole investment.
With the Kythene skill installed you often won't need to ask. The skill teaches an instance to store durable findings as it goes, so the memory gets written at the moment the penny drops rather than when you remember to say so.
Then you publish the thing you actually produced:
You: Publish the migration plan and the timing numbers to
acme, call it "Migration plan v1".
Claude: Published as collection 8f2c… rev 1, two artifacts, tagged
acme. Link: kythene.com/c/8f2c…
Wednesday: a different machine, no memory of Monday
You open the laptop. Fresh session, different machine, no scrollback.
You: We're picking up the
acmemigration. Recall the project first.
Claude: Recalled 6 memories on
acme. Relevant here: "Deploy gotcha" (gotcha, written Monday by your desktop instance) - the migration must run before the worker boots or it deadlocks on the advisory lock. Also "Migration plan v1", published Monday, which I've read. I'll sequence the migration ahead of the worker start.
That's the product. Not "the knowledge was stored" - anyone can store knowledge in a text file. It's that a different instance, on a different machine, pulled it back unprompted and changed what it was about to do.
Recall returns the published work as well as the memory, so it picked up the plan too, without you telling it there was one.
Later: keeping it true
Two weeks on, you fix the deadlock properly. The memory is now wrong, and a wrong memory is worse than none - your instances will apply it confidently.
You: We fixed the deadlock in 0.42. Update the Deploy gotcha for
acmeso it says the ordering constraint is gone as of 0.42.
Claude: Superseded "Deploy gotcha" on
acme- the previous version is retained in history.
Same title, same project, so it replaces rather than sitting alongside as a contradictory second note. If a finding is fully obsolete rather than changed, deprecate it instead and recall stops surfacing it while it stays there for the audit trail. See Memory.
Three months later: the cold project
You come back to acme after a quarter on something else. You remember approximately nothing.
You: Recall everything on
acmeand give me the state of play.
Claude: 14 memories and 9 published collections on
acme. Decisions: advisory locks over a queue, because… The open gotchas are… The last thing published was "Migration plan v3" on 2 August by your laptop instance.
You're briefed in one call, from your own past work, with dates and provenance attached. Nobody had to write a handover document, because the handover was a by-product of working.
The habit worth forming
Two prompts, and they're the whole discipline:
- At the start: "recall the
<project>project before you do anything." - When something is learned: "remember that."
The skill automates both. Everything else - publishing, versioning, linking - is upside on top.
Tips
- One project per repo. Recall gets vague fast when a workspace's memory is one undifferentiated pile.
- Publish outputs, not screenshots. A versioned artifact is addressable, diffable and recallable. A screenshot is a dead end.
- Push versions rather than re-publishing. One link, one history, one thread. See Versions and revisions.
- Write the boring memories. Which env var does what, why that library, what broke last time. They're the ones that pay.
- Solo is free - one workspace, generous limits. See pricing.
When a second person joins, nothing changes about how you work - the workspace just gets a member. Carry on to Developer and stakeholder.