Cortex — local second brain (Chrome extension)

Cortex extracts readable page text (Mozilla Readability), chunks it for retrieval (not one blob per page), stores rows in IndexedDB (Dexie), and embeds chunks locally with Transformers.js (Xenova/all-MiniLM-L6-v2). Search blends semantic similarity, keyword/title signals, recency, and a simple importance score from visits — without sending data to a backend (there is none).

Indexing aims at meaningful text the user encounters, with pause/blocklist controls — not “record everything forever.”

Build

cd cortex
npm install
npm run build
npm run test

CI

Workflow .github/workflows/ci.yml runs npm ci, typecheck, tests (including eval), and production build with working-directory: cortex. If your Git repo root is the extension directory only, remove that working-directory block and fix cache-dependency-path.

Load cortex/dist via chrome://extensions → Developer mode → Load unpacked.

Toolbar icons

1. Put your square brand PNG at cortex/icons/cortex-brand.png (the halftone mark).
2. Run npm run icons — this uses Sharp to write icons/icon-16.png, icon-48.png, icon-128.png.
3. manifest.json already points action.default_icon and icons at those paths (copied to dist/icons on build).

First semantic indexing downloads ONNX/model weights (~22 MB + wasm/caches) from Hugging Face / jsDelivr — manifest.json includes those hosts unless you bundle weights under vendor/models/ (see vendor/models/README.md).

Usage

• Browse normally — after idle time per navigation, readable text is chunked and queued for embeddings (service worker + offscreen document).
• Cmd + Shift + K (Mac) / Ctrl + Shift + K (Windows/Linux), or Alt + Shift + C as an alternate — toggles the in-page overlay. If nothing fires, fix bindings under chrome://extensions/shortcuts. The extension also injects content.js the first time you use the shortcut on a tab that never ran Cortex (e.g. opened before install); internal Chrome URLs (chrome://, PDF viewer, etc.) cannot be scripted.
• Ask in plain language — e.g. who works at Afterquery, works in Afterquery yesterday, or "Afterquery" with yesterday / today / last week. Parsing boosts company-like phrases, optionally prefers LinkedIn profiles, and can filter by visit-log time. Results are snippets from your index with sources (extractive recall), not cloud ChatGPT. Popup shows documents, chunks, visit log size, recent visits, and privacy toggles (pause indexing, domain blocklist).

Privacy / limits

• No screenshots, no keylogging — extracted text only (same class of data as Readability).
• Incognito tabs are not indexed by default (guard in the service worker).
• Heuristic sensitive-site skip (banking/health/gov patterns, etc.) reduces accidental capture; refine with the blocklist.
• Heavy ML runs in an offscreen document (offscreen.js ~750 KB minified); the service worker stays smaller.

SPA sites (LinkedIn, etc.)

Cortex intercepts pushState / replaceState / popstate, debounces MutationObserver updates, and schedules retries so dynamic pages can be captured after content loads.

Pipeline (MVP)

Content script → extract → chunk → background queues embeddings → IndexedDB (documents + chunks) → hybrid search → overlay.

Deleted GhostWriter

The previous ghostwriter folder could not be removed automatically because Windows reported it “in use”. Close anything locking it (Chrome holding unpacked extension, terminals, editors), then delete ghostwriter manually if it still exists — cortex is the replacement project under project/cortex.