filed under: the re-paste tax
Context that writes itself.
A CLAUDE.md is a letter to your AI that you have to keep rewriting. Custom instructions are the same letter, copied into a different mailbox. Both were true the day you wrote them. And every day since, a little less. Working Memory skips the authorship: it fills itself in as you work.
14 days free · no credit card required
written by hand vs written by working
The context file, side by side.
| Dimension | context files | working memory |
|---|---|---|
| Written by | You. By hand. Usually on a Sunday. | Itself, as you work. |
| Kept current by | You, remembering to edit it. | The work itself. New decisions file as you make them. |
| Where it applies | One repo, or one app's settings box. | Every AI you've connected. |
| When a new AI joins | Paste. Adapt. Repeat. | Sign in, click Allow. |
| What each prompt carries | The whole file, every time. | Just what's relevant, recalled on demand. |
| Control | Total: versioned, reviewable, in git. | Editable, exportable, deletable. Curated as you go. |
skills teach. memory knows.
Not a skill. A memory.
Worth clearing up, because they rhyme: a skill is a context file with ambition. You write instructions (a procedure, a format, a workflow) and your AI reads them to learn how to do something. Authored by hand, and stale the same way.
Memory runs on the other axis. Not how to do things, but what's true: your projects, your people, your decisions, the why behind them. Working Memory isn't a skill you install or a file you keep current; it's a connector your AI plugs into over MCP, and it fills itself in as you work. Skills you write. Memory, you shouldn't have to.
we keep one ourselves
When a CLAUDE.md is plenty.
Team conventions belong in a context file. Build commands, code style, "never edit a shipped migration": anything everyone working in a repo, human or AI, must read belongs in git, versioned and reviewed. We keep a CLAUDE.md in our own repo, and this page was written next to it.
What a context file is wrong for is you. Your preferences, your decisions, the context that follows you across projects, apps, and machines. Files stay in the repo they serve. You don't.
(Keep both. The file holds the team's rules; Working Memory holds yours.)
retire the paste buffer
Stop re-introducing yourself.
Connect once per AI, then just work. Your context stays current because it's collected while it's happening, not the last time you remembered to edit a file.
Connect your AI14 days free · no credit card required
elsewhere in this series · vs built-in ai memory · vs building your own