Bear Brain

A second brain for Bear notes, powered by Simon Willison’s llm.
It syncs your Bear notes into a local SQLite DB, chunks them intelligently by headings, embeds them, and lets you query them with any LLM supported by llm.

---

Setup

1. Install dependencies

pip install -r requirements.txt
llm install llm-sentence-transformers

2. Initialize database + first sync

./setup.sh

This creates brain.db, copies/normalizes Bear’s DB, and runs the first sync.

⸻

Syncing Notes

To keep up to date:

./bear-sync.sh

• Reads Bear’s DB (safe, read-only snapshot)
• Diffs against the last sync timestamp
• Splits changed notes into heading-aware overlapping chunks
• Embeds new/updated chunks into the bear_chunks collection in brain.db

Run manually, via a cron/launchd job, or trigger before queries.

⸻

Asking Questions

./brain_ask.sh "what was that idea about curriculum scheduling?"

Options:

• -m → pick a specific model (e.g. one hosted in LM Studio or OpenAI)
• Uses --no-stream by default for reliability
• Pulls top-K similar chunks, adds them as context, and asks the LLM to answer with [ID] citations

⸻

How It Works

• Chunking: Notes are split on Markdown headers (#, ##, etc.) with overlapping windows (~1200 chars, 200 overlap). Each chunk has a stable chunk_id.
• Embeddings: Stored with content (--store) so retrieval returns text directly.
• Incremental sync: Only re-embeds notes that changed since the last run.
• Collection: All embeddings live in the bear_chunks collection.

⸻

Status

Work in progress, tested only on macOS. Use at your own risk.