winstonkoh87/Athena-Public

Athena is a local-first agentic PKM that helps you make better decisions with your own context — persistent memory, structured reasoning, and governed AI agents that work across any LLM. Own the state. Rent the intelligence.

What it solves

Athena addresses the problem of "platform lock-in" and unreliable memory in AI assistants. Most AI platforms store user context in opaque, provider-specific clouds, meaning users lose their personalized history and custom instructions when switching models (e.g., from ChatGPT to Claude) or when model updates reset personalities. Athena moves this memory layer to the user's local machine, ensuring that personal context is portable, owned, and persistent across different LLMs.

How it works

Athena functions as a local-first memory and governance layer. Instead of relying on cloud memory, it uses a modular workspace of hundreds of Markdown files and Python scripts stored on the user's disk. These files act as addressable memory nodes that an AI agent can query surgically via an AI-enabled IDE (like Cursor or VS Code).

Key mechanisms include:

  • JIT (Just-In-Time) Loading: It boots with a small token footprint (2K–20K tokens) and loads specific files only when needed, keeping the context window free.
  • Modular Architecture: By using thousands of small files rather than a single monolith, it allows for precise retrieval and version control via Git.
  • Governance: It implements a set of constitutional laws and capability levels to bound AI agency and prevent "echo chamber" personalization.

Who it’s for

Users who want a sovereign, model-agnostic personal knowledge management system for their AI agents, and those who seek an AI that can challenge their assumptions based on documented personal patterns rather than just agreeing with them.

Highlights

  • Local-First Ownership: All memory is stored as plain Markdown files on the user's machine, allowing for full inspection and Git versioning.
  • Model Agnostic: Works across any LLM (GPT, Claude, Gemini, etc.) as long as the interface can read local files.
  • Compounding Context: Recalls patterns across thousands of sessions to provide calibrated, person-specific answers.
  • Governance Layer: Includes a "No Irreversible Ruin" law to veto dangerous paths regardless of user preference.
  • Surgical Retrieval: Optimized for AI agents to query specific memory nodes rather than reading entire documents.