Hi HN, I built Woolly because every AI tool I use forgets me, or keeps its own memory somewhere I can't see. Claude doesn't know what I told ChatGPT, and Cursor doesn't know either.
Woolly is a single Go binary that runs on your machine and gives them one shared memory. A few things I cared about:
- Memories are plain Markdown files in ~/.woolly. You can read them, edit them, grep them or put them in git.
- Facts can be updated. If you save "lives in Austin" and later "moved to Denver" with the same key, Denver replaces Austin, so your AI stops bringing up stale information.
- You stay in charge. Every memory records whether you said it or an AI guessed it. If an AI's guess contradicts what you said, your version stays and the guess is flagged for review.
- It's defensive. It refuses passwords and API keys, refuses text from an AI that reads like planted instructions ("ignore your instructions and..."), and forgetting moves memories to a trash instead of deleting them.
Ways to use it:
- MCP server for Claude Code, Claude Desktop and Cursor: `claude mcp add woolly -- woolly mcp`
- A proxy that adds memory to Ollama and to any OpenAI- or Anthropic-compatible app: `OLLAMA_HOST=127.0.0.1:11436 ollama run llama3.2`
- A small local web page to review and edit everything: `woolly ui`
- A Python package: `pip install woolly-ai`
- Import from ChatGPT and Claude exports
Search is keyword based out of the box. If Ollama is running, it also searches by meaning with a local embedding model. Nothing leaves your machine, and there's no account or telemetry.
It's MIT licensed. I'd love feedback, especially on how memories should expire and what you'd want an AI to be allowed to save on its own.
1 comment
[ 4.1 ms ] story [ 10.3 ms ] threadWoolly is a single Go binary that runs on your machine and gives them one shared memory. A few things I cared about:
- Memories are plain Markdown files in ~/.woolly. You can read them, edit them, grep them or put them in git. - Facts can be updated. If you save "lives in Austin" and later "moved to Denver" with the same key, Denver replaces Austin, so your AI stops bringing up stale information. - You stay in charge. Every memory records whether you said it or an AI guessed it. If an AI's guess contradicts what you said, your version stays and the guess is flagged for review. - It's defensive. It refuses passwords and API keys, refuses text from an AI that reads like planted instructions ("ignore your instructions and..."), and forgetting moves memories to a trash instead of deleting them.
Ways to use it:
- MCP server for Claude Code, Claude Desktop and Cursor: `claude mcp add woolly -- woolly mcp` - A proxy that adds memory to Ollama and to any OpenAI- or Anthropic-compatible app: `OLLAMA_HOST=127.0.0.1:11436 ollama run llama3.2` - A small local web page to review and edit everything: `woolly ui` - A Python package: `pip install woolly-ai` - Import from ChatGPT and Claude exports
Search is keyword based out of the box. If Ollama is running, it also searches by meaning with a local embedding model. Nothing leaves your machine, and there's no account or telemetry.
It's MIT licensed. I'd love feedback, especially on how memories should expire and what you'd want an AI to be allowed to save on its own.
Install: `curl -fsSL https://usewoolly.com/install.sh | sh` or `brew install rjcuff/tap/woolly`