Project Athena

A local-first agentic PKM that helps you make better decisions with your own context.

Capture what you learn, retrieve it when relevant, and carry your context between models — a local-first memory, reasoning, and governance layer for any LLM, across ChatGPT, Claude, Gemini, and whatever you switch to next.

Own the state. Rent the intelligence. Platforms forget. Athena doesn't.

Most assistants remember you so they can agree with you faster. Athena remembers you so it can tell you when you're wrong.

Quickstart · How It Works · Validation Status · Docs · FAQ · Safety · Contributing

Last updated: 27 July 2026

The Problem

You've spent months training ChatGPT to understand you. Then a model update resets the personality. Your custom instructions stop working. You can't find that conversation from last Tuesday. And if you switch to Claude or Gemini? You start from zero.

Platform memory is unreliable, opaque, and locked to one provider. You don't own it, you can't inspect it, and you can't take it with you.

Why Athena?

Athena moves the memory layer to your machine. Plain Markdown files that you own, version-control, and point at any model.

• 🧠 Your Memory, Your Machine — Files on your disk, not in OpenAI's cloud. Read them, edit them, git-version them.
• 🔌 Switch Models Freely — Claude today, Gemini tomorrow, GPT next week. The memory stays. The model is just whoever's on shift.
• 📈 It Compounds — Session 500 recalls patterns from session 5. The durable asset isn't the code — it's your data: anyone can fork Athena; nobody can fork your sessions. That's a personal continuity advantage, not vendor lock-in — the files move with you. Honest caveat: compounding needs curation. Keep the /end loop running; unpruned memory decays like any archive. → The Compounding Effect
• ⚡ 2K–20K Token Boot — Scales to the task. Lightweight chat (~2K) → /start (~10K) → /ultrastart (~20K). 80–98% of your context window stays free, even after 10,000 sessions.
• 🔬 Meta-Game Reasoning — Generic LLMs optimise within the game you're playing. Athena asks whether you should be playing that game at all. → Meta-Game Thesis
• 🛡️ Governed Autonomy — 6 constitutional laws, 4 capability levels, bounded agency.

A generic LLM is a brilliant amnesiac. Athena is the hippocampus — the memory that makes intelligence useful.

Or in engineering terms: The LLM is the engine. Athena is the chassis, the memory, and the rules of the road. Swap the engine anytime — the car remembers every road you've driven.

The design philosophy: augment the human, not replace them. After 1,900+ sessions, the bottleneck shifted — optimising the operator is now higher-leverage than optimising the AI.

The Human Augmentation Thesis

Athena's centralised design principle: augment human cognition, not replace it. The more context you give Athena, the sharper its answers become — not by remembering your preferences, but by reasoning differently because of what it knows about you.

But personalization is only half the design — and on its own, it's the dangerous half. An AI tuned purely to fit you is a mirror: it hands your own blind spots back to you, faster and more fluently than you'd rationalise them yourself. The moat was never that Athena agrees with you more precisely. It's that Athena knows you well enough to tell you when you are the problem — and has the standing (Law #1, the Committee of Seats) to refuse a premise a generic assistant would obligingly help you execute.

So the USP has two legs, not one:

• It knows you — owned, portable, compounding context. The substrate nobody else has.
• It will disagree with you — grounded in your documented patterns, not generic hedging. The leg that makes the first one safe to stand on.

Personalization is what makes the disagreement credible (it's aimed at your actual situation, not a textbook). The disagreement is what keeps the personalization from becoming a well-decorated echo chamber. Neither leg is the USP alone; the product is the pair.

A generic LLM gives the internet's statistically average answer — correct on average, across all humans. Athena gives answers calibrated to your specific situation — including the situations where the honest answer is the one you were hoping it wouldn't say:

Generic LLMs solve the question. Athena solves the person. The same question, asked by different people with different lives, demands fundamentally different answers. A generic LLM can’t differentiate because it has no context. Athena can’t give the same answer twice — because the context files are different. The memory is the product — and what the memory buys is the right to push back. Solving the person isn’t flattering the person: the same context that personalises the answer is what licenses Athena to hand you the one you were hoping to avoid.

What Athena Actually Does With Your Problem

Not all problems are solvable. Athena classifies and responds accordingly:

The uncertainty of the domain changes Athena's conviction level: