A model on your own GPU reads the page and writes a few sentences. Only those sentences reach your memory layer.
Not on the Chrome Web Store yet. Two-minute install below.
Lists your spaces so you pick one.
Lists the users it already holds.
Re-remembering updates, never duplicates.
Adding another is one file. No telemetry, and no magpie server. Your key goes straight from the browser to whichever you chose.
All of it. That is the whole thing that gets sent.
270× the summary, one square each.
Summary is the default. Page text is opt-in, and providers bill on what they extract.
git clone https://github.com/anonalabs/magpie
npm install && npm run vendor:wasm && npm run build
chrome://extensions → Developer mode → Load unpacked → pick dist/
Open the popup, choose your memory layer, paste an API key.
Alt Shift M
The first capture downloads the model once, then it is cached.
| Model | Video memory | Context | For |
|---|---|---|---|
| Gemma 3 1B | 711 MB | 4,096 | integrated graphics |
| Qwen 2.5 1.5B | 1,630 MB | 4,096 | a good default |
| Llama 3.2 3B | 2,264 MB | 4,096 | best summaries |
Also reads PDFs, and keeps a highlighted passage exactly as you selected it.