🌐 mnemosyne-os.io — the product · mnemosyne-os.com — for organizations · 📖 docs.mnemosyne-os.io — the documentation

Lab Watch

The new publications of the AI labs, every day, in one window.

The name is provisional: no trademark search has been done.

Lab Watch is a cartridge for Mnemosyne OS. It hands the host's background watch six targets: one OpenAlex list per lab (OpenAI, Anthropic, Google DeepMind, Meta, Mistral AI) and the OpenAI news feed. The host checks them once a day while Mnemosyne OS runs, window closed or not. When you open Lab Watch, it lists what arrived since you last looked, by lab, with the publication date, the day it was picked up and a link.

Design and measurements: docs/architecture/187_research-crossing-lab-publications.md.

Sources

Source	What it gives	Read by
OpenAlex, one list per lab	articles and preprints, Zenodo deposits left out, newest 50	the host's watch (json mode), then the frame for the details
openai.com/news/rss.xml	every OpenAI post; research categories listed, the rest folded	the host's watch (atom mode), then social.fetch

• A work belongs to a lab when one of its authors does, as OpenAlex attributes it. The card says so.
• The first pass of the watch announces nothing.
• An abstract under 20 words is marked too short to read.

Your topics

Write a few topics in your words. On a click, the app's model reads each new abstract and says which topic it touches, quoting a sentence of the abstract word for word, or that it touches none. The number of model calls is on the button; the credits spent are measured on the balance afterwards. Verdicts are kept, so reopening the window costs nothing.

The lot 0 measure of this judge (103 lab publications labelled by hand) is ready behind ?bench=lot0 on the frame address. It has not been run.

Cross a text

Load a lab's publications (one OpenAlex credit per 200), paste or pick your text (txt, md, pdf, docx). Each passage is shown the five publications closest by their words and judged: supported, partly, diverges (both quotes side by side, yours to judge) or absent, which always says out of how many. A publication with an open PDF can be read in full; its passages then join the comparison.

Permissions

• watch:background: the host keeps checking while the window is closed.
• vault:read: the gate of social.fetch (the OpenAI feed, no CORS header) and of reader.downloadUrl / reader.extractDocument (open PDFs, your PDF or Word file). Lab Watch reads no vault.
• model:infer: the judge.
• wallet:read: the credit balance, to measure what a reading cost.
• dialog:open: picking and reading your file.

Run it

pnpm dev     # vite on 127.0.0.1:5246
pnpm test
pnpm typecheck

Outside Mnemosyne OS the window says it has no host. A local stand-in that answers like the host lives in .preview/host.html (gitignored).

Where Mnemosyne OS lives

This cartridge runs inside Mnemosyne OS, the sovereign, local-first memory operating system published by XPACEGEMS LLC. Its official addresses:

• Product site: https://mnemosyne-os.io
• Organizations: https://mnemosyne-os.com
• Documentation: https://docs.mnemosyne-os.io
• Host source: https://github.com/Mnemosyne-OS/Mnemosyne-Neural-OS
• Packages: the npm scope @mnemosyne_os

---

_{Mnemosyne OS — the sovereign, local-first memory OS this cartridge runs in. Get it at mnemosyne-os.io/download, install cartridges from the built-in MnemoHub store, or build your own.}