Verbs and outcomes
The one with the 5/5 cleanup pilot score.
HDC Note Codebook
Turn 10,000 markdown files into a six-codebook hyperdimensional index that runs on a laptop. The same library we use internally to make Claude remember our notes from last quarter. CPU-only, ~20 MB on disk per 10K notes, MIT licensed.
Why we built it. Off-the-shelf embeddings are slow and memory-hungry on big personal corpora. Hyperdimensional computing gives you sub-second cleanup retrieval on a CPU at 1–5 KB per note. We use the same library to drive every retainer client’s private assistant.
Index 10K notes, ask a question.
# clone and install
$ git clone https://github.com/squatch-c-c/hdc-note-codebook
$ cd hdc-note-codebook && pip install -e .
# index a folder of markdown
$ codebook index ~/notes --out notes.codebook
[ok] scanned 9,438 files
[ok] built 6 codebooks (action / entity / topic / time / cite / mood)
[ok] wrote notes.codebook (19.2 MB)
# ask a question
$ codebook ask notes.codebook "what did I write about retainers in March?"
→ retainer-pricing-v2.md 0.84
→ march-leads.md 0.71
→ tom-jerick-call.md 0.62Works on Python 3.10+. Pure NumPy. Optionally integrates with Claude Code, Cursor and Continue.dev via the included MCP server.
Six codebooks, one CLI.
People, products, projects
Who and what a note is actually about.
Conceptual buckets
Subject matter, independent of wording.
Rolling temporal index
“Last quarter” resolves to actual files.
Quotes, links, sources
What a note borrowed, and from where.
Sentiment / voice fingerprint
How it was written, not just what it says.
All six pack into one ~20 MB file per 10K notes and load in milliseconds. Update incrementally — no full re-index when you add a file.
Three signs you need this.
Your notes folder is big
More than ~2,000 files, and grep stopped being an answer a while ago.
Embeddings got expensive
You tried embedding-based search and it eats RAM, or fees, or both.
The notes stay home
You want an assistant that reads your private notes without sending them anywhere.
We integrate this in week one.
If you want the codebook running on your corpus, behind a private chat UI, with eval and ops dashboards — that’s a pilot or a retainer. We scope it on a thirty-minute call.
Just want release notes? Ask to be added — added by hand, no list software.