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.

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 persona you build, then call by name — the wizard asks what it is called before it asks anything else, then pronouns, the language it replies in and how it writes. You open with "Julia, look at this" the way you would with a person, it answers as that name rather than as "the assistant", and it is the same character on every surface. Renaming it later is a line in ~/.config/agent-julia/config.json plus agent-julia sync.

• A voice you shape by talking — "stop hedging", "don't use that word", "write to my co-founder, not to an engineer". Say it once, in passing, and it holds. Withdraw it the same way.

• One memory, every surface — Claude Code and Claude Desktop share the same brain; a fact saved in one is known in the other.

• You own it — plain markdown in a git repo you control, not a database locked inside an app. Readable, portable, git log-able, and no write can silently destroy what it was meant to extend.

• Small context footprint — only a budgeted persona core rides each turn; everything else is recalled on demand, so the window stays clear.

Related MCP server: LLM Memory MCP Server

Install

npx agent-julia init

Requires Node.js 24+ (it uses the built-in node:sqlite — no native module to compile or rebuild). There's no separate install step: npx fetches agent-julia into its own cache and runs the wizard in one go, and the MCP server launches the same way (npx -y agent-julia@latest serve), so it stays current. Prefer it installed permanently? npm i -g agent-julia.

The wizard starts with the name you will be calling it by, then pronouns, language, and voice; where your memory lives and whether to version it with git; search; and which Claude apps to register. It either writes a small persona block into each app's startup context — backed up first, inside a marked block, fully reversible — or prints the exact changes for you to make by hand (see Manual setup).

When you're done, restart your Claude apps so they pick up the new MCP server.

Usage

You don't call tools or memorize commands. You talk to your agent by name, and it reaches into memory on its own. A few natural examples:

Save something (→ ingest)

• "Remember that we dropped Redis — we're on Postgres LISTEN/NOTIFY now."

• "Note for the Privé project: the weekly review moved to Mondays."

• The agent also tries to capture durable facts on its own — but treat that as a bonus, not a guarantee (see the note below). Saying "remember: …" is the reliable channel.

On automatic capture. agent-julia is an MCP server, which is pull-only: it never writes on its own, only when the model calls a tool. So "save this" happens only if the agent decides to call ingest mid-conversation — reliable when you ask explicitly ("remember: …"), best-effort when left to the agent's judgment, and easy to miss during a focused task. The persona core nudges it to capture and to say what it saved, so a miss is visible. If something important didn't land, just say "remember: …".

Recall (→ search, read)

• "What did we decide about auth?"

• "What do you know about the Privé pricing model?"

• "Did I say anything about onboarding last week?"

Steer the voice (→ correct_voice)

• "Stop hedging — commit to an answer."

• "Don't open every reply with my name."

• The correction is recorded and applied from the next turn, on every surface.

Browse (→ list)

• "What's in my memory?" · "List every page you have."

Think something through (→ the brainstorm skill)

• "Help me decide whether to take the Berlin offer."

• "Let's brainstorm the pricing for the new plan."

• "Back to the office-move question from last week." — sessions are resumable; the state lives in memory, so you can start on one machine and finish on another.

Everything you save lands as markdown in your repo and is committed, so you can open it, edit it by hand, or git log the history any time.

Why it helps

Two everyday frustrations it removes:

A preference you keep re-teaching. You tell Claude Code to skip the "Great question!" preamble and reply in Polish.

• Before — tomorrow, in Cowork on your laptop, the preamble is back and it answers in English. You correct it again. And again.

• After — you record it once with correct_voice; Code and Cowork both read the same correction from the next turn on.

The CLAUDE.md tax. Your ~/.claude/CLAUDE.md has grown to 500 lines — every project, every preference, every person.

• Before — all of it loads on every turn of every session, and the answer degrades as the window fills with things this task doesn't need.

• After — a persona core sized by contextBudget (1,200 tokens by default, plus a short instruction telling the agent where its memory lives) rides each turn; the 500 lines live in your memory, and only the page you actually need is pulled in when you ask.

How it works

• Canonical store — plain markdown in a git repo. Human-readable, portable, versioned, private. This is the source of truth, not a database.

• Derived index — SQLite (FTS5 full-text + optional vector embeddings) built from the markdown, via Node's built-in node:sqlite (no native module to compile). It's disposable: delete it and it rebuilds itself from your files.

• Budgeted core — a compact persona block is injected into Claude Code's CLAUDE.md and Claude Desktop's global instructions. It stays within a token budget you set, so it never crowds out the conversation. Each layer is budgeted separately, and the part that yields is ours: identity and the privacy rail come off the top, your voice and your corrections each get a share, and the shipped universal rules take what's left. maintenance and doctor both say when something didn't fit, rather than trimming in silence.

• One MCP server — every surface talks to the same agent-julia server over stdio, so they share one memory and one persona.

Skills

agent-julia ships one skill, installed into ~/.claude/skills/ (which both Claude Code and Cowork read) by the wizard and refreshed automatically every time the server starts — so existing installs pick it up on their next session, no manual step:

brainstorm — a structured facilitator that takes a fuzzy idea or an open decision to a committed outcome. Built for business and life questions, not code. What makes it different from just chatting:

• A real process — frame (question, criteria, constraints) → diverge → challenge (assumptions, pre-mortem, inversion) → converge on a recommendation you confirm section by section.

• One question per message, with answer options — a dialogue, not a questionnaire.

• Thinking lenses — distillations of proven tools (strategy kernel, jobs-to-be-done, willingness to pay, fear-setting, inversion, regret minimization, one-way vs two-way doors…), picked to fit the topic and named when used.

• Deep mode — for high-stakes calls the agent offers a panel of 3–4 adversarial perspectives (skeptic, the other side, operator…) run as parallel subagents where available. It costs real tokens, so it never runs without your yes.

• It remembers — session state is saved to memory after every phase, so a brainstorm survives days and devices ("back to the X question"); the outcome lands as a decision doc with the options you killed and why, so nothing gets relitigated.

• Your places, not hardcoded ones — at the end it offers the doc as a Markdown/HTML file and, based on what tools you have connected, proposes where the actions should go.

The skill starts only when you ask for it (or accept a one-line offer). Your own skill under the same name is never overwritten: installs and uninstalls only touch copies carrying the author: agent-julia marker.

Beyond Claude

The persona you built is not locked to Claude. agent-julia export codex writes it into ~/.codex/AGENTS.md, export gemini into ~/.gemini/GEMINI.md, and export <path> into any AGENTS.md-style instruction file — as a marked, removable block, backed up on first touch, and kept refreshed by the server on every boot (a voice correction you make in Claude reaches Codex on its next session). The export carries the persona itself — identity, voice, corrections, privacy rail — without the Claude-specific memory instructions; attach the same MCP server in the other tool if it supports MCP, and the memory follows too.

Reference

Deeper detail — search, persona, the file layout, and the full configuration and command reference.

Search

Two layers work together, and both run locally.

Keyword (always on). SQLite FTS5 with Porter stemming, so debug matches debugging, and diacritics folding, so cafe matches café and lodz matches Łódź (the letters SQLite cannot fold on its own — ł, đ, ø, ß — are carried by a folded shadow column, so titles and snippets keep their real spelling). For languages without spaces between words — Chinese, Japanese, Korean, Thai — Agent Julia switches to a trigram tokenizer, with a substring fallback for the one- and two-character queries that are ordinary in those languages. A question asked as a sentence is answered by a ladder: all terms, then content words only, then any two, then the most distinctive one alone. Each hit says which rung it came from, so a loose answer looks loose.

Meaning (optional). Turn on semantic search to find a note even when you phrase it differently, and across languages — a question in Polish can surface an English note. You choose how it runs:

• Local model — a multilingual model (the multilingual-e5 family, ~118 languages) runs in-process. Pages are embedded in parts, split on their own headings, so a long page is searchable to its end and not just to the model's first 512 tokens. No server, no API key, fully offline after a one-time model download. Pick a size in the wizard: small (~120 MB download, ~0.3 GB RAM), base (~280 MB, ~0.6 GB), or large (~560 MB, ~1.3 GB). The size is a one-time download cached on disk; the model also loads into RAM while search runs. The wizard reads your machine's RAM and CPU cores and suggests a tier. RAM is rarely the limit — even the largest model needs only ~1.3 GB — so the suggestion leans on cores, since a bigger model's main cost is slower CPU inference per query. The real trade-off is download size and speed against quality. Needs one extra package, @huggingface/transformers, which stays optional so the base install is tiny.

• Hosted API — any OpenAI-compatible endpoint (OpenAI, or a local server like Ollama or LM Studio). Your key is read from an environment variable and never written to disk.

• None — stay keyword-only. The default, and completely dependency-free.

Hybrid mode blends keyword and meaning, and degrades gracefully: with no embeddings configured, it's simply keyword search.

Persona

The persona has three layers, with a clear order when they disagree:

1. User corrections (highest) — short notes you record over time, like "don't use the word X" or "less hedging". Captured with the correct_voice tool, kept in voice-corrections.md, and applied above everything else. It's your agent.

2. Universal core — a small set of communication rules that apply to every persona: talk like a person, lead with a recommendation, skip filler.

3. Style (lowest) — either one of four shipped voices (sharp co-founder, calm mentor, minimalist engineer, neutral assistant) or your own voice. For the presets, the wizard shows the same message in all four, in your language, so you choose by ear. For a custom voice, pick "Write my own voice" and describe how the agent should speak in persona.md; it's used in place of a preset.

The composed persona core — identity, universal core, your style or custom voice, and corrections — is injected into each surface's startup context within your token budget, so the agent is always in character without fetching anything. (The get_core tool returns the same core on demand.)

Name, pronouns, and language are yours to set. The default persona is "Julia" (she/her), but you define your own.

Memory model

Your store follows a simple, enforced layout:

your-memory/
  index.md     catalog of pages (kept current automatically)
  log.md       append-only journal of changes
  pages/       one page per topic, kebab-case
  archive/     retired pages, read-only

Pages carry light frontmatter — title, status, last-updated date, and an auto-detected language — and link to each other with [[wiki-links]]. Writing always goes through the ingest tool, which writes the page, refreshes the catalog, appends the journal, reindexes, and commits to git in one step. A write either appends to the page or replaces it, says which it did and by how much, and refuses a replace that would throw an established page away. agent-julia undo walks the commits it wrote and reverses one.

Point the wizard at an existing markdown knowledge base and Agent Julia adopts it: your pages are indexed and a hand-written index.md is left alone — Agent Julia only manages a clearly marked block inside it.

When a project documents itself

Most projects already document themselves, and that documentation should not become a copy inside your store: it is large, it changes daily, and a copy goes stale the same week. A page can say where the rest of what it is about actually lives, and Agent Julia routes there instead.

project: ~/Sites/prive
sources:
  - kind: dir
    at: _doc
    about: product and technical documentation
  - kind: mcp
    at: acme
    how: docs_search, docs_read
    about: company documentation, reachable without the repo

kind is dir, file, mcp or url, so a documentation tree, a company MCP server, a Notion space and a repo's own CLAUDE.md are all declared the same way. A string works too ("dir:_doc — the product docs"). A documentation directory and a CLAUDE.md in the project are found without being declared; a server or a URL has to be written down, because no directory listing reveals it.

The route then travels wherever the question is asked: search attaches it to a hit on that page, read appends it to the page, and the block in Claude Code's own memory directory names it for an agent standing inside the repo. Every one of them carries the same rule — ask Agent Julia about the user, their projects and past decisions, read the project's own documentation for how the project works.

Claude Code's own memory

Claude Code keeps a memory of its own: one directory per working directory under ~/.claude/projects/<project>/memory, an index it puts into the prompt of every session started there, and one file per fact. It does the same job as this store and reaches exactly one directory — a fact saved there is invisible to Claude Desktop, to Cowork, and to the same repo opened from another machine.

Agent Julia settles it. codeMemory in the config takes three values:

Directories are never created: only the ones Claude Code made for itself are touched, the block goes in next to whatever you wrote there, and every file replaced by a pointer is backed up first. agent-julia uninstall removes the blocks and restores the files.

A directory holding more than 25 facts of its own is left alone even under absorb, and reported instead: that much is a knowledge base someone chose to keep next to the repo, not a handful of strays, and folding it in unasked would bury a decision you made. agent-julia sync --absorb-all brings it in when you mean to.

How the persona reaches each surface

Two halves, delivered differently, because one of them can be written for you and the other cannot.

The server also hands every client a short set of instructions when it connects — who the agent is, what it must never store, and the instruction to load the rest. That part needs no paste at all, so a machine where you never finished the setup still gets a named agent that knows what it must not keep.

agent-julia doctor reports what it actually knows here, and says plainly what it cannot: it has no way to read the in-app field, so it tells you what it last asked you to paste, what the last Desktop session was seen running with, and whether the voice has genuinely been fetched there.

Optionally back the store with a git remote — a private GitHub repo, say — set in the wizard or later with agent-julia remote <url>. By default it syncs on maintenance (and server startup), best-effort, so an offline moment or a missing credential never blocks a write; it just pushes on the next run. Turn on gitAutoPush to push after every write instead, trading a network round-trip per write for immediate off-machine backup. agent-julia push syncs on demand. With a remote set, the server also pulls on startup — so on a second machine, a session starts from what the first one pushed; a merge conflict is aborted (never left half-done) and reported for a by-hand agent-julia pull.

You drive Agent Julia by talking to it (see Usage); these are the underlying tools and the CLI.

MCP tools (the agent calls these for you):

The persona core is also exposed as a resource (agent-julia://core) for clients that prefer resources to a tool call.

CLI commands (you run these):

Settings live in ~/.config/agent-julia/config.json and carry a schemaVersion. The wizard writes it for you; the fields:

The wizard can write the Claude config changes for you, or print the exact edits so you make them yourself — it asks which you prefer, and never touches a Claude file without that choice. To see the steps any time:

agent-julia sync --print

What it adds:

Claude Code — in ~/.claude.json, add the server under mcpServers:

{ "mcpServers": { "agent-julia": { "command": "npx", "args": ["-y", "agent-julia@latest", "serve"] } } }

Then append the persona block (printed by the command) to ~/.claude/CLAUDE.md.

Claude Desktop — add the same mcpServers entry to the Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS), and paste the short persona block into Settings → Instructions for Claude. That field lives inside the app and no program can write it, so it is the one manual step — which is exactly why what goes in it is only the part that never changes: the name, the pronouns, the reply language and the never-store list. Your voice and your corrections are not in it; the agent fetches those at the start of a conversation, so they are current without you pasting again. agent-julia paste prints the block and copies it.

Dispatch (mobile) isn't supported. It can't run or reach the local stdio MCP server, and it doesn't read Cowork's Global instructions — so neither the memory tools nor the persona reach it in a local-first setup. Covering Dispatch would need the server hosted remotely, which is out of scope here.

The persona block is small on purpose — identity plus an instruction to use agent-julia for memory. Everything else is fetched on demand. When the wizard does this for you, it backs up each file first, writes inside a marked block, and agent-julia uninstall removes it cleanly.

Housekeeping runs on its own. On every write, and again when the server starts, Agent Julia reindexes changed pages, picks up files you edited by hand, flags stale-dated notes and broken links, refreshes the catalog, recompacts the persona core, and commits. Nothing is deleted without you — stale items are flagged, not removed.

A heavier weekly pass — merging duplicates, retiring stale pages, fixing broken links — is owner's-judgment work, so it runs as a digest: ask your agent to "run the weekly digest" (or schedule a recurring Cowork task with that prompt), and it gathers the candidates — near-duplicate pages (by embedding similarity), long-untouched facts, orphan [[links]], unlinked pages, oversized pages — and walks you through them one at a time. Nothing changes without your yes; approved retirements go to archive/ (kept on disk and in git history, out of the index). agent-julia maintenance runs the automatic half from the terminal.

Upgrades are backward-compatible, or they ship an automatic migration that runs on first launch — backed up, idempotent, and transparent. The config carries a schemaVersion; ordered migration steps bring older stores forward on startup. The derived search index is disposable and simply rebuilds itself when its shape changes. @latest picks up new versions automatically next session; pin an exact version (agent-julia@<version>) if you want it fixed. Either way, upgrades never lose data and never ask you to hand-edit files.

Releasing (maintainers)

CI runs typecheck, build, and tests on every pull request and push to main (Node 24). Releases are tag-driven:

# add the new section to CHANGELOG.md, then:
npm version <patch|minor|major>
git push --follow-tags

A vX.Y.Z tag triggers the release workflow: it checks the tag against package.json, runs the tests, publishes to npm via Trusted Publishing (OIDC — no stored token, provenance attached automatically), and cuts a GitHub release from the matching CHANGELOG section.

Credits

The persona's "avoid the tells of AI writing" rules are adapted from the humanizer skill by Siqi Chen (MIT), itself based on Wikipedia's "Signs of AI writing" (WikiProject AI Cleanup, CC BY-SA).

License

MIT.