Project Athena

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.

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. Platform memory decays; Athena's doesn't. The moat isn't the code — it's your data. Anyone can fork Athena; nobody can fork your sessions. → 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 1,500+ sessions, the bottleneck shifted — optimising the operator is now higher-leverage than optimising the AI.

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.

A generic LLM gives the internet's statistically average answer — correct on average, across all humans. Athena gives answers calibrated to your specific situation, because statistical correctness and personal correctness are completely different things:

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.

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

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

As uncertainty increases, Athena shifts from "here's the answer" to "here's the valid structural zone" to "here are your options — you choose." This is deliberate: false confidence in stochastic domains is more dangerous than honest uncertainty. Athena's conviction is proportional to domain determinism and context completeness.

Crucially, conviction and decisiveness are independent axes. Low certainty about outcomes doesn't require vague output. A surgeon operates with high decisiveness and low conviction about outcomes. In semi-stochastic domains, Athena delivers precise, operational setups — then explicitly defers the probability judgment to you. "Setup: Long 1.0850 / SL 1.0800 / TP1 1.0920. Your calibration: structural tell present Y/N?" — not "you might want to consider..." — Protocol 524 →

Law #0 (Sovereignty): Your life, your weights, your choice. Law #1 (No Irreversible Ruin): …unless the choice ends the game permanently. Law #1 overrides Law #0. Always.

Athena doesn't tell you what you should do. It shows you what you can do, what each option costs, and hands the choice back. The only exception: paths that end the game permanently.

Architecture, not oracle. This domain classification is a replicable architecture — each Athena instance calibrates independently over time through bilateral use. Session 1 treats most problems conservatively. Session 500 has accumulated enough frameworks, case studies, and corrected assumptions to tighten confidence bands and solve more sub-problems autonomously. The calibration compounds; the model is interchangeable. — Protocol 525 (Cross-Domain Weighting) →

Kind of. But first, it helps to understand what those names actually refer to — because they blend three very different things:

When people say "ChatGPT remembers me," they mean the platform stores some memory on their cloud. When they say "Claude is smart," they mean the model reasons well. When they say "Cursor writes code," they mean the IDE connects model + files.

Athena is none of these. Athena is the memory and governance layer that sits between your IDE and whichever model you choose — owned by you, stored on your machine, portable across all three layers.

There's a difference between remembering your name and thinking in your frameworks:

💡 Think of platform memory like photos on Instagram — you can view them, but you don't own them, can't move them, and can't search them precisely. Athena is keeping the originals on your hard drive, with albums, metadata, and full edit history.

Works on macOS, Windows, and Linux.

git clone https://github.com/winstonkoh87/Athena-Public.git
cd Athena-Public

Clone it anywhere you keep projects (e.g. ~/Projects/). This folder is your Athena workspace — your memory, protocols, and config all live here.

# Create and activate a virtual environment
python3 -m venv .venv
source .venv/bin/activate   # macOS / Linux
# .venv\Scripts\activate    # Windows

# Lightweight install (~30 seconds, no ML dependencies)
pip install -e .

# Full install (~5–10 min, enables vector search and reranking)
pip install -e ".[full]"

⚠️ Don't pip install athena-cli — that's a different, unrelated package. Always install from inside the cloned repo.

Open the Athena-Public/ directory as your workspace root in one of these editors:

• Antigravity · Cursor · Claude Code · VS Code + Copilot · Kilo Code · Gemini CLI

In your IDE's AI chat panel (e.g. Cmd+L in Cursor, the chat sidebar in Antigravity), type:

/start

/tutorial

This walks you through everything: what Athena is, how it works, builds your profile, and demos the tools (~20 min). Confident users can skip it.

/end

That's it. No API keys. No database setup. The folder is the product.

Every session follows one cycle. Three modes let you match overhead to task complexity:

flowchart LR
    subgraph "🟢 Lightweight Mode"
        A1["Chat"] --> C1["/end"]
        C1 --> D1["Memory"]
    end

    subgraph "🔴 Full Boot Mode"
        A2["/start"] --> B2["Work"]
        B2 --> C2["/end"]
        C2 --> D2["Memory"]
        D2 --> A2
    end

    subgraph "⚫ Deep Boot Mode"
        A3["/ultrastart"] --> B3["Work"]
        B3 --> C3["/ultraend"]
        C3 --> D3["Memory"]
        D3 --> A3
    end

    style A1 fill:#3b82f6,color:#fff,stroke:#333
    style C1 fill:#ef4444,color:#fff,stroke:#333
    style D1 fill:#8b5cf6,color:#fff,stroke:#333
    style A2 fill:#22c55e,color:#fff,stroke:#333
    style C2 fill:#ef4444,color:#fff,stroke:#333
    style D2 fill:#8b5cf6,color:#fff,stroke:#333
    style B2 fill:#3b82f6,color:#fff,stroke:#333
    style A3 fill:#1e1e1e,color:#fff,stroke:#555
    style B3 fill:#3b82f6,color:#fff,stroke:#333
    style C3 fill:#ef4444,color:#fff,stroke:#333
    style D3 fill:#8b5cf6,color:#fff,stroke:#333

Why this happens: The AI model doesn't improve — your data does. Each /end extracts decisions, patterns, and learnings into your memory bank. The next /start loads that accumulated intelligence. Same algorithm, better data, better output. → The Compounding Effect

Athena is modelled after the human body. Built bottom-up by the creator. Used top-down by the user.

"As within, so without, as above, so below." — Same pattern at every layer. Fractal by design.

Everything you need to turn a generic AI into your AI — pre-configured, no assembly required.

Athena works through AI-enabled code editors — apps that connect to AI models while reading your local files. It does not work through ChatGPT.com, Claude.ai, or Gemini web — those are closed sandboxes that can't read your disk.

More agents planned — full compatibility list →

"How is this different from ChatGPT Projects?" — Projects reset every new chat and are locked to one platform. Athena persists across all chats, all models, with full version history. Details →

📖 Deep Dive: How Athena Solves — The Three Core Use Cases — covers the vulnerability prerequisite, pre-work convergence, domain reclassification, the EEV decision framework, privacy architecture, and honest limitations.

The asymmetry. A licensed therapist costs $200+/hr. A business coach costs $500+/hr. A negotiation consultant costs $1,000+/hr. Athena gives you structured, context-aware guidance across all of these domains — 24/7, for the cost of your existing AI subscription. It doesn't replace professionals for clinical emergencies, but for the 90% of life decisions and psychological patterns that don't require a medical license, it closes the gap between having access to wisdom and not being able to afford it.

⚠️ Important: Athena is an experimental AI tool, not a licensed professional service. It cannot diagnose, treat, or manage any medical or psychiatric condition. See SAFETY.md for crisis contacts and responsible use guidelines.

Not just for coding. Athena is used for personal knowledge management, health tracking, creative writing, business strategy, and daily life — by people who've never written a line of code.

Athena is free. Forever. MIT licensed. You only pay for the AI subscription you're probably already paying for.

Try before you buy. Athena works with Google Antigravity's free tier — clone the repo, type /start, and see if it clicks. No credit card, no trial period, no catch. Upgrade only when you hit the free tier's daily limits.

Why $200/mo sounds expensive — until you do the math. A single employee costs $5,000+/mo in salary, benefits, and management overhead. An AI agent on a max-tier subscription costs $200/mo, works 24/7, doesn't call in sick, and scales to any domain you throw at it. For best results, subscribe to the max plan on any one platform (Claude Max, Google AI Ultra, etc.) — the difference between $20/mo and $200/mo is the difference between a tool you use occasionally and a tool that runs your life. Heavy users routinely consume $2K–$3K+ in equivalent API costs per month — the flat subscription turns variable cost into fixed cost.

For peak performance, use /ultrastart every session. On a flat-rate plan, the marginal cost of deeper thinking is $0 — so the cost of under-thinking always exceeds the cost of over-thinking. /ultrastart loads ~20K tokens of structured context (identity, canonical memory, active state, semantic bridge) every session. On API pricing, this costs $2–5/session. On a $200/mo flat plan, it costs nothing. The pricing model of the underlying compute layer directly determines Athena's performance ceiling.

Boot cost is 2K–20K tokens (depending on mode) — constant whether it's session 1 or session 10,000. Details →

Athena-Public/
├── athena.yaml              # Agent manifest — model, tools, hooks, governance
├── src/athena/              # SDK package (pip install -e .)
│   ├── core/                #   Config, governance, permissions, security
│   ├── tools/               #   Search, agentic search, reranker, heartbeat
│   ├── memory/              #   Vector DB, delta sync, schema
│   ├── boot/                #   Orchestrator, loaders, shutdown
│   ├── cli/                 #   init, save, doctor commands
│   └── mcp_server.py        #   MCP Tool Server (9 tools, 2 resources)
├── tools/                   # Declarative tool definitions (YAML)
├── scripts/                 # Operational scripts (boot, shutdown, launch)
├── examples/
│   ├── protocols/           # 150+ starter frameworks (15 categories)
│   ├── scripts/             # 540+ reference scripts
│   └── templates/           # Starter templates (framework, memory bank)
├── docs/                    # Architecture, benchmarks, security, guides
└── pyproject.toml           # Modern packaging

🌟 Star History

MIT License · Contributing · Safety · Security · Code of Conduct

Clone it. Boot it. Make it yours.