Agent Julia — AI skill for Claude Code

Agent Julia

**Give your AI a name, a voice, and a memory that stays.** You start by building the persona: what it is called, its pronouns, the language it answers in, how it writes, what it must never store. From then on you address it by that name instead of prompting an assistant, and the same character answers on Claude Code and on Claude Desktop. It is yours: plain markdown in a git repo you own, kept small in the model's context, corrected in passing and remembered for good.

[](https://www.npmjs.com/package/agent-julia) [](./LICENSE)

npx agent-julia init

Every Claude surface forgets between sessions and keeps its own separate notes, so the same person gets a slightly different, amnesiac assistant on each one — one that has to be re-taught the same preferences and called nothing in particular. The usual fix, piling everything into `CLAUDE.md` and global instructions, backfires: long context measurably degrades the model, and a pasted block goes stale the moment you correct something.

Agent Julia gives your AI an identity instead of a configuration. The first thing the wizard asks is what to call it, and from then on that is who answers: you write "Julia, take a look at this" and Julia replies, in her own voice, on Claude Code and on Claude Desktop alike. You build that voice by talking, correcting it the way you would correct a colleague ("stop hedging", "don't open with my name"), and it keeps the correction everywhere from the next turn on. Its memory is one folder of markdown you own, with a small budgeted slice of persona riding each message and the rest recalled only when it is needed.

Highlights

• **A pers