An AI chief of staff that reads your stuff, remembers what matters, and runs your life in the background. Forks in 30 minutes. Runs on your existing Claude subscription.
In 1998, transhumanists coined exocortex to describe a hypothetical information-processing system that behaves as if it were part of your brain — extending memory, attention, and judgment outside your skull. For 25 years it was a thought experiment. Then Claude shipped agents that could read your messages, hold persistent context, and act on a schedule. Now it's a weekend project.
This repo is the scaffold. Clone it, fill in seven memory files with the texture of your actual life, point Claude at it, and you have an always-on operator who runs a daily briefing, tracks every promise you've made, and tells you what to do next — with one-tap "I did this" links so the loop closes even when you go offline.
Most "AI assistants" are reactive. You open them, you type, they answer, you close them, nothing persists. That's a tool, not a brain extension.
An exocortex has three properties:
- It observes. Reads your inbound — WhatsApp, Gmail, Calendar, Tasks, SMS — without you having to ask.
- It remembers. Persistent state across sessions, in plain markdown you can edit by hand.
- It acts on a schedule. Runs even when you're asleep. The morning briefing is on your phone before your alarm.
This repo wires all three.
Seven scheduled tasks already wired up:
- 5am daily — Morning briefing. Scans your WhatsApp groups + DMs, Gmail, Calendar, and tasks. Flags commitments, unanswered threads, and what's on your plate today. Emails you the result with one-tap acknowledgment links for every action item.
- 9pm daily — Evening reflection. Catches plans confirmed after the morning pass, updates memory.
- Sunday 9am — Weekly review. Prunes stale memory (3-week rule), refreshes priorities, reviews relationship health.
- 7am daily — Commitment → Calendar. Turns every "let's grab coffee Saturday" into a real calendar event.
- Friday noon — GSD agenda. Builds a punch list from your tasks + open commitments for a Friday focus block.
- Friday noon — Weekend Fun Finder. Searches local events, scored against your household's preferences.
- 28th monthly — Financial sweep. Spending autopsy, net worth update, subscription audit, credit card optimization.
Five on-demand skills:
/draft— Writes outbound messages in your voice. Studies the last 10–20 messages in any thread before drafting./reply— Three reply options + a recommendation for any message thread./wildcard— One contrarian, non-obvious insight on whatever's on your mind./gsd— Prioritized punch list, right now./roast— Group-chat roast on demand. Tune it to your circle's tone.
Most personal-AI projects fail in predictable ways. Here's what we did differently.
1. Persistent memory in plain markdown. Seven files. Claude reads them at the start of every task, updates them at the end. No vector DB. No RAG. No fine-tuning. Just files. The reason this works: markdown is human-editable, version-controllable, and grep-able. You can correct your exocortex's memory by opening a file in any editor. You own the state. (See docs/memory-system.md.)
2. The acknowledgment loop. This is the hardest problem in personal AI: how do you close the loop on actions that happen offline? You called your mom. There's no API event. The agent has no way to know.
Solution: every briefing email includes one-tap mailto: links next to action items. Tapping one fires a self-email with a coded subject — EXO_DONE::call-mom — that the next briefing parses. Match the slug, mark the commitment done, archive the signal. (See docs/acknowledgment-system.md.)
3. Tier-weighted attention. Not all messages are equal. The morning sweep applies five tiers (Family → Extended family → Close friends → Network → Other), so a "hey when are you free?" from your partner surfaces before a meme from a group chat. The tier system is in CLAUDE.md — change it to match your life. (See docs/customization.md.)
Three layers. Full breakdown in ARCHITECTURE.md.
┌──────────────────────────────────────────────────────────┐
│ Layer 1 — The Briefing Engine (scheduled tasks) │
│ morning · evening · weekly · daily commitment sweep · │
│ GSD agenda · weekend finder · monthly finance │
├──────────────────────────────────────────────────────────┤
│ Layer 2 — The Memory System (seven markdown files) │
│ people · commitments · life · insights · │
│ finances · health · carry-on (private, gated) │
├──────────────────────────────────────────────────────────┤
│ Layer 3 — Skills (on-demand commands) │
│ /draft · /reply · /wildcard · /gsd · /roast │
└──────────────────────────────────────────────────────────┘
▲
│ reads/writes via MCP connectors
▼
┌──────────────────────────────────────────────────────────┐
│ WhatsApp · Gmail · Google Calendar · Google Tasks · │
│ Android Messages · Fitbit · Monarch Money │
└──────────────────────────────────────────────────────────┘
Full walkthrough in SETUP.md. The 30-second version:
- Install Claude Desktop with Cowork mode enabled.
- Clone this repo somewhere stable on your machine.
- Mount the folder in Cowork.
- Open
CLAUDE.mdand fill in the placeholders ({{USER_NAME}},{{PARTNER_NAME}}, your tier system, etc.). - Connect the MCPs you want — WhatsApp, Gmail, Calendar are the high-leverage trio.
- Tell Claude: "Read CLAUDE.md and onboard me. Walk me through filling in the memory files."
- Once memory is bootstrapped, tell Claude: "Set up the seven scheduled tasks per BACKLOG.md."
The first morning briefing lands at 5am the next day.
Three knobs to turn:
Persona. Edit the Persona section of CLAUDE.md. The default is "direct, warm, slightly cheeky chief of staff who tells you what you need to hear, not what you want to hear." If you want softer or more sarcastic, change it there. Every output inherits.
Tier system. The default tiers are placeholder categories (Family / Meaningful work / Health). Yours might be Cofounders / Investors / Side project. Whatever they are, write them down — the exocortex filters everything through this lens.
Scheduled tasks. Each briefing in briefings/ is a standalone prompt. Edit them, remove the ones you don't want, add new ones for your situation. The founder version probably wants a daily "what moved the needle?" reflection. The new-parent version probably wants a school-pickup buffer alert.
- Claude Pro or Max subscription: $20–$200/mo. Max recommended if you're running all seven scheduled tasks against a busy life.
- MCP connectors: free for personal accounts (Gmail, Calendar, WhatsApp, Fitbit). Monarch Money is $15/mo if you want financial sweeps.
- Your time: ~3 hours upfront to bootstrap memory honestly. ~5 min/week to review what the exocortex surfaces. The 3 hours pay back inside a month.
You will get value out of this if you:
- Run a complicated life (family + work + side projects + community + investing) and lose threads regularly.
- Already use Claude or ChatGPT a lot and feel the pain of starting from scratch every conversation.
- Are willing to write down what actually matters to you in markdown.
You will not get value if you:
- Want a chatbot you talk to. The exocortex runs in the background; you barely interact with it.
- Are uncomfortable letting an LLM read your messages and email. It does. It has to.
- Want zero configuration. The configuration is the product — your context is what makes it work.
- All memory lives on your machine. Nothing leaves your laptop except Claude API calls and the briefing emails you send to yourself.
- The
memory/private/carry-on.mdfile is strictly gated — never read automatically, never referenced in any output, only loaded when you explicitly invoke it. Use it for anything sensitive. .gitignoreexcludesmemory/from version control by default. Your second brain stays out of GitHub.- Read
docs/privacy.mdbefore connecting anything financial or medical.
exocortex/
├── README.md ← you are here
├── ARCHITECTURE.md ← the three-layer system, in depth
├── SETUP.md ← step-by-step onboarding
├── BACKLOG.md ← what to build next (yours to fork)
├── CLAUDE.md ← the persona + ruleset (fill in placeholders)
├── memory/ ← your second brain (gitignored)
│ ├── people.md.template
│ ├── commitments.md.template
│ ├── life.md.template
│ ├── insights.md.template
│ ├── finances.md.template
│ ├── health.md.template
│ └── private/carry-on.md.template
├── briefings/ ← the seven scheduled task prompts
├── skills/ ← /draft, /reply, /wildcard, /gsd, /roast
├── docs/ ← deep dives on tricky bits
└── examples/ ← a sanitized example morning briefing
If this is useful, star the repo — it helps others find it. If you've extended it in interesting ways (new briefing types, new connectors, new skills), open a PR. I'd especially love to see versions tuned for founders, students, parents of young kids, and retirees.
MIT. Take it, fork it, sell it, whatever. Just don't claim you invented the acknowledgment loop without crediting it — that one took a while to figure out.