Iantha

Personal chief of staff + knowledge base curator, in markdown. A clone-and-run repo for Claude Code.

Iantha is a memory + skills bundle that turns Claude Code into a personal assistant for daily life — tasks, reminders, decisions, routines, anything you'd otherwise keep in your head. Clone the repo, launch Claude Code inside it, and you have an assistant that remembers across sessions.

Optionally point Iantha at an Obsidian vault and it doubles as a knowledge-base curator — daily notes, decisions, weekly reviews, and a reading list, all in plain markdown. The pattern is inspired by Andrej Karpathy's LLM Knowledge Bases framing: raw material lands in your vault, an LLM helps you compile it into something queryable, and you keep editing in your IDE of choice (Obsidian).

⭐ Star this repo if you're running Iantha — it helps others find it.

Why

Claude Code remembers nothing between sessions. Mention "remind me Friday" or "I've decided to stop X" — next session starts blank. There's no structure for tasks, routines, or the life context you'd actually want an assistant to know.

Iantha is that structure: memory files auto-captured from chat, a /setup onboarding interview plus six daily-rhythm skills (/morning, /evening, /debrief, /obsidian, /housekeep, /consolidate-learning), and an optional Obsidian vault. The part you'd build yourself after a month of using Claude Code for life admin.

Setup

Your memory is personal — tasks, people, routines — so it belongs in your own private repo, not a public fork. Two ways in:

Use it for real (recommended). Click Use this template → create a private repo → clone it:

git clone https://github.com/<you>/<your-iantha>.git ~/iantha
cd ~/iantha
claude

Now origin is your private repo: commit your memory and it follows you across machines, private to you. (Want skill updates later? git remote add upstream https://github.com/kiloloop/iantha.git and merge when you like.)

Just trying it. A plain clone is fine for a test drive — but origin points at this public repo, which you can't push to, so re-point it at your own private repo before relying on cross-machine sync:

git clone https://github.com/kiloloop/iantha.git ~/iantha
cd ~/iantha
claude

Either way, Iantha reads CLAUDE.md and the memory/ directory at session start. To make it yours in one sitting, type /setup — a short, skippable interview that seeds your memory with your real tasks, priorities, routines, and people, then runs /morning on it. Prefer to skip it? Just start talking — Iantha learns as you go, and /setup is re-runnable later to fill in the gaps.

First setup: use the template, run claude, type /setup — a few minutes. Every day after: just /morning.

What's Inside

.
├── CLAUDE.md              # Persona, rules, memory map
├── README.md              # You're reading it
├── config.yaml            # Optional vault config (vault_dir + subdirs)
├── .claude/skills/        # /setup, /morning, /evening, /debrief, /obsidian, /housekeep, /consolidate-learning
├── memory/                # Tasks, personal context, decisions, learnings
└── vault-template/        # Starter VAULT.md to copy into your Obsidian vault

Daily Use

Command	When	What it does
/setup	first run (or to fill gaps)	Short onboarding interview — seeds your memory with your real life, then briefs you
/morning	start of day	Briefs you on today's tasks, priorities, anything overdue
/evening	end of day	Archives what got done, rolls recurring tasks, notes tomorrow
/debrief	end of substantive session	Captures decisions, learnings, feedback into memory
/obsidian	as needed	Vault read/write primitives (no-op without vault_dir)
/housekeep	weekly	Audits memory + skills + vault hygiene
/consolidate-learning	monthly	Promotes skill "Learned from runs" entries into main instructions

Plus two more skills installed from kiloloop/oacp-skills (see "Shared Skills" below) — /self-improve, /wrap-up.

Outside those, just talk to Iantha. Mention "tomorrow" or "by Friday" — it auto-captures. Share preferences or routines — it remembers. Ask for context — it surfaces what's relevant.

Examples

These match the sample seeds the repo ships with (memory/*.md.example) — your first /morning on a fresh clone demos exactly this, labeled as sample data, while your real memory files start blank. The sample persona is a tech lead with a side project about to ship and a life around it — swap in your own.

Auto-capture from chat. Mention something time-bound in normal conversation and Iantha files it:

You: I need to call mom tomorrow at 6pm.

Iantha: Added to tasks — call Mom, tomorrow 18:00. I'll surface it in tomorrow's /morning briefing.

Decision logging in passing. Mention a decision and Iantha writes it to memory/decisions.md so it doesn't evaporate:

You: I've decided to switch the team standup from daily to weekly. Signal-to-noise was too low.

Iantha: Logged in memory/decisions.md with today's date, rationale captured. Want me to add updating the recurring invite to tasks?

Memory

Iantha's persistent state is in memory/. The live files start blank and fill in as you use it; the *.md.example files carry sample seeds (same convention as .env.example) that power the first /morning demo — say "delete the samples" once you're running on your own data.

File	Purpose
tasks.md	All actionable items (urgent / work / personal / recurring / archive)
personal.md	About you — preferences, routines, people, life context
priorities.md	What you're focused on right now
decisions.md	Decisions you've made, with date and rationale
feedback.md	Corrections you've given Iantha (so it doesn't repeat mistakes)
learnings.md	Operational knowledge — tool gotchas, workflow insights
MEMORY.md	Thin index, loaded at session start

The repo is git-tracked. Commit your memory updates and your context follows you across machines.

Trust & Privacy

Iantha is files on your disk, nothing else.

• Nothing leaves your machine beyond your normal runtime traffic to its model provider (Claude Code → Anthropic, Codex → OpenAI) — no extra services, no telemetry, no third-party calls.
• Memory is plain markdown. Every fact Iantha knows about you is a diffable file in memory/ you can read, edit, or delete — and git log is the audit trail of what changed and when.
• No self-granted authority. Memory is treated as executable policy: Iantha's rules forbid storing anything that would grant itself approval or identity shortcuts (the failure shape Anthropic documents in the Fable 5 system card, §2.3.3) — if such an entry ever appears, it gets flagged to you instead of followed.
• The vault is optional and just as local. Leave vault_dir unset and Iantha never touches Obsidian.

Knowledge Base mode (optional)

Set vault_dir in config.yaml to point Iantha at an Obsidian vault. Then:

• /morning and /evening append to ${vault_dir}/daily/YYYY-MM-DD.md
• /obsidian decision files decisions into ${vault_dir}/decisions/
• /obsidian reading-list queues docs for your review
• /obsidian lint flags broken wikilinks, stale _to-delete/ items, orphans

Run /obsidian init once to seed the vault with the standard directories and copy vault-template/VAULT.md into it. The handbook there is the contract Iantha follows when writing.

# config.yaml
vault_dir: ~/Documents/MyVault

If vault_dir is unset, every vault hook no-ops silently — Iantha works fully without a vault.

Shared Skills

Two skills come from kiloloop/oacp-skills (catalog: https://oacp.dev/skills/) rather than being shipped in-repo. Install once with:

npx skills add kiloloop/oacp-skills self-improve wrap-up

You get:

• /self-improve — audit CLAUDE.md, skills, and memory for drift
• /wrap-up — end-of-session orchestrator (cleanup + debrief + self-improve + commit). Composes with the in-repo /debrief.

Optional, install per-need: /doctor, /check-inbox, /review-loop-*.

Multi-runtime setup (optional)

If you run Iantha alongside other agents (Claude Code + Codex + Gemini, etc.) and want them to share session memory across runtimes, look at kiloloop/cortex — a cross-session memory layer (SSOT + debrief inbox) built on the OACP protocol. Cortex publishes its own /debrief and /sync skills wired to a shared inbox; Iantha's in-repo /debrief is the single-agent equivalent.

For solo personal use, you don't need cortex.

Using with Codex / other runtimes

Iantha has no runtime lock-in — it ships ready for both Claude Code and Codex from the same clone. Nothing here is proprietary; the two runtimes just look in differently-named places, so the repo provides both, pointing at one set of files:

	Claude Code reads	Codex reads	How it's shared
Instructions	CLAUDE.md	AGENTS.md	AGENTS.md is a symlink to CLAUDE.md
Skills	.claude/skills/	.agents/skills/	.agents/skills/ is a symlink to .claude/skills/
Memory + config	memory/, config.yaml	same	shared as-is — not runtime-specific

So on Codex you clone and go, same as Claude Code: launch Codex in the repo and it loads its persona/rules from AGENTS.md and discovers the skills under .agents/skills/ — the very same files Claude Code uses. The two helper scripts (/obsidian, /housekeep) are pure Python that runs anywhere.

Customizing carries across both — edit CLAUDE.md to rename the persona or tweak a rule, or edit a .claude/skills/<name>/SKILL.md, and both runtimes follow, because the symlinks point at those originals.

Notes:

• If your platform doesn't preserve git symlinks (some Windows setups), recreate them from the repo root: ln -s CLAUDE.md AGENTS.md and ln -s ../.claude/skills .agents/skills.
• The same shared-file pattern extends to any other AGENTS.md-style runtime (Gemini, etc.) — add a symlink at the path that runtime expects.

Personalize

The name "Iantha" is a default. To rename:

1. Edit CLAUDE.md — replace Iantha with your preferred name
2. Edit README.md and memory/MEMORY.md headers
3. Commit

Memory file conventions (timestamps, sections) are defined in memory/MEMORY.md. Adjust to your style.

For machine-specific overrides (real vault paths, your name) without polluting the public repo, create config.local.yaml next to config.yaml — it's gitignored and merges over the base config.

Roadmap → v0.2.0

The current release covers the chief-of-staff half well. The knowledge-base half is intentionally minimal — Karpathy-style wiki compilation comes in v0.2.0:

• /obsidian compile — compile raw/ (clippings, transcripts, raw notes) into curated wiki/ entries
• /obsidian ask — Q&A over wiki/ from the CLI
• Web-clipper / read-it-later ingest into raw/
• Marp slide rendering for wiki entries
• Vault search (full-text + semantic)

If you start collecting raw material in raw/ now, the v0.2.0 compiler will work on what's already there.

Upgrade Path

Iantha is the free, self-hosted starter. If you want a managed multi-agent version that coordinates across projects (work + life, multiple agents, cloud-synced), look at Iris — the paid product. Iantha is your starter Iris.

Why "Iantha"

Greek for "violet flower." Iantha is the flower-pair sister of Iris (the rainbow goddess and flower) — naming chosen because the free starter complements the paid Iris.

License

Apache 2.0.