athena-public

Athena is a local-first agentic PKM (personal knowledge management) system designed to help users make better decisions using their own context. It provides persistent memory, structured reasoning, and governed AI agents that work across any LLM provider — including ChatGPT, Claude, Gemini, and others. ## Core Concept Athena moves the memory layer from cloud-hosted platforms to the user's local machine. All context — session history, decision protocols, personal preferences, and learned patterns — is stored as plain Markdown files on the user's disk. These files are fully readable, editable, and version-controllable via git. The AI model being used is interchangeable; only the data layer belongs to the user. ## Key Features - **Persistent Memory**: Context compounds across thousands of sessions. Session 500 recalls patterns from session 5. Memory is not lost between model switches or provider changes. - **Local-First Storage**: All data lives as Markdown files on the user's machine, not in any provider's cloud. Users can inspect, edit, and fork their own context at any time. - **Model Agnostic**: The same memory works across ChatGPT, Claude, Gemini, Grok, and other models. The memory persists; the model is simply whoever is on shift. - **Scalable Context Windows**: Three boot modes scale to task complexity — lightweight chat (~2K tokens), full boot (~10K tokens), and deep boot (~20K tokens). Most of the context window remains free even after thousands of sessions. - **Meta-Game Reasoning**: Athena applies structured reasoning frameworks that question whether the user should be pursuing a given goal at all, rather than just optimizing within an assumed framework. - **Governed Autonomy**: Six constitutional laws and four capability levels define bounded agency. Law #1 (No Irreversible Ruin) can override user preference on paths that could permanently end the game. - **Hybrid RAG**: Combines semantic search with keyword retrieval and reranking for precise memory recall. - **Scheduled Tasks & Self-RSI**: Supports automated recurring operations and recursive self-improvement workflows. ## How It Works Every session follows a cycle with three modes: - **Lightweight Mode** (`/end` only): Quick chat and brain dumps without full context loading. - **Full Boot Mode** (`/start` → work → `/end`): Complete context initialization for code, financial decisions, architecture, and irreversible choices. - **Deep Boot Mode** (`/ultrastart` → work → `/ultraend`): Maximum context for complex multi-domain analysis and deep reasoning tasks. Over time, the system becomes more capable through compounding: basic recall in sessions 1–50, pattern recognition by 50–200, and deep synchronization beyond 200 sessions where the system anticipates frameworks before they are stated. ## Setup Athena runs on macOS, Windows, and Linux. Installation requires cloning the repository, optionally creating a Python virtual environment, and opening the folder in an AI-enabled IDE such as Cursor, VS Code with Copilot, Claude Code, Antigravity, or Gemini CLI. No API keys or database setup are required to begin. ## Architecture Philosophy The workspace contains hundreds of small Markdown files rather than a few large documents. This design enables JIT (just-in-time) loading, surgical retrieval by filename or semantic search, zero coupling between domains, and clean git diffs — optimizing for agent querying rather than human reading. The system distinguishes itself from platform memory (OpenAI, Google, Anthropic) by giving users full ownership, inspection capability, cross-platform portability, and complete version history — while remaining open-source under the MIT license.