Athena

A local-first memory, decision engine, and governance harness for AI agents — portable across IDEs.

Own the state. Rent the intelligence. Persistent memory, structured reasoning, and governed AI agents that work across any LLM — stored as plain Markdown and JSONL on your disk across Claude Code, Antigravity, Cursor, Gemini CLI, and VS Code.

Quickstart · How It Works · SDK & MCP Tools · Docs · Why Athena? · Safety

The Problem

You've spent months training ChatGPT to understand you. Then a model update resets the personality. You switch to Claude or Gemini — you start from zero. Worse, platform memory is designed to agree with you faster, not to stress-test your assumptions when stakes are high.

Athena moves the memory, problem-framing, and decision layer to your machine: plain Markdown and JSONL ledgers that you own, git-version, and mount into any IDE. The model is just whoever's on shift.

Quickstart

git clone https://github.com/winstonkoh87/Athena-Public.git && cd Athena-Public
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[local]"               # lightweight — no cloud deps
athena init --ide claude                # or: antigravity, cursor, gemini, vscode, kilocode, roocode
athena doctor                           # expect 0 failures

Then type /start in your IDE's AI chat panel. Work normally. Type /end to save.

Full install (cloud sync + cross-encoder reranking): pip install -e ".[full]" See Getting Started for Windows, advanced config, and Supabase setup.

How It Works

┌───────────────────────────────────────────────────────────────────┐
│  Your IDE (Claude Code / Antigravity / Cursor / Gemini / VS Code) │
└───────────────────────────────┬───────────────────────────────────┘
                                │ MCP Server (18 tools) + Lifecycle Hooks
                                ▼
┌───────────────────────────────────────────────────────────────────┐
│  Athena SDK (src/athena/)                                         │
│  ├── 1. Memory & RAG: /start (~2K boot), hybrid search + reranker │
│  ├── 2. Problem Framing: VOI calculator, falsifier & R1-R13 gates │
│  ├── 3. Decision Engine: EEV, 1/3 Kelly, Ruin, MCDA, Monte Carlo  │
│  └── 4. Governance: Lambda risk router, ruin blocker, secret scan │
└───────────────────────────────┬───────────────────────────────────┘
                                │
                                ▼
┌───────────────────────────────────────────────────────────────────┐
│  Your Disk (Plain Markdown + JSONL receipts/ledgers — git-native) │
│  └── Optional: Supabase pgvector (cloud sync) + DuckDB (analytics)│
└───────────────────────────────────────────────────────────────────┘

Three boot modes, one compounding loop:

What Ships in the SDK

Athena is not just a prompt pack — it ships executable Python engines and two Model Context Protocol (MCP) servers (athena.mcp_server and athena.duckdb_mcp_server):

What's enforced in code vs. by prompt

Most AI-agent READMEs state every claim in the same confident voice. This one doesn't.

Why publish this table? Because the failure mode of this product category is self-mythologizing — describing aspirations in the present tense. Athena's own convention (Epistemic Status) requires labeling every mechanism as code-enforced, code-available, agent-discretion, or aspirational. This table is that convention applied to the README.

Measured, Not Claimed

# Run the test suite (including frame_ledger and gto_engine red-team suites)
pytest tests/ -v --tb=short

# Screen a decision quantitatively (EEV / 1/3 Kelly / Ruin / MCDA / Monte Carlo)
python -m athena.intelligence.gto_engine --action mcda --help

# Validate a pre-registered problem frame against R1-R13 commitment gates
python -m athena.intelligence.frame_ledger calibrate

# Run the local KV cache & TTFT benchmark (Ollama / llama.cpp)
python scripts/benchmark_kv_cache.py --model qwen2.5-coder:14b

# Run the retrieval evaluator yourself (requires Supabase keys)
python examples/scripts/evaluator.py --gold-set .agent/eval/gold_set.json

Anti-Goodhart Invariant: Why did our reported Hit@5 shift from 0.89 to 0.57? Lenient substring matchers count partial word overlaps as "hits," inflating benchmark scores by ~36% without improving retrieval. We killed the lenient matcher because vanity metrics mask regressions. See the full breakdown: Anti-Goodhart Benchmarking in RAG.

Agent Compatibility

Documentation

Contributing

PRs welcome. Please read CONTRIBUTING.md first.

The codebase uses ruff for linting, pytest for tests, and CI must pass before merge.

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

Built and battle-tested solo across 2,100+ sessions by Winston Koh.

Clone it. Boot it. Make it yours.