MIT ~2,000 lines of TypeScript · Read all of it ↗

Your memory never leaves.
Nothing is forgotten

A private memory engine for your Mac. It remembers what you read, copy, save, and screenshot — and hands that context to any AI you talk to. No cloud. No subscription. Yours.

Install Memex Star on GitHub
Claude Memex connected ·
MEMEX VAULT
what was that essay about doing great work I read last week?
Found it in your vault — "How to Do Great Work" by Paul Graham. You read it Tuesday morning. The core argument: pick something you have a natural aptitude for and an obsessive interest in, then work on your own projects at the frontier.
🌐How to Do Great Work · paulgraham.com · visited Aug 11, 9:09 AM
summarize the part about curiosity
Memex · searching your vault…

Every conversation starts
with amnesia

Your assistant knows nothing about the forty articles you read, the PDFs you saved, the links you copied. Memex runs quietly in the background, indexes it all into a local database, and serves it to every AI you use. Your memory, on tap.

Private by architecture,
not by promise

Zero cloud. Your data lives in one SQLite file in your home directory. Period.
Only the links you copy. OCR runs on Apple's on-device Vision framework, search is local, and there is no analytics of any kind. The one thing Memex fetches is a URL you copied, so it can archive the page. On first start it archives up to 500 recent clipboard links. Verify with the network tab.
Zero accounts. No signup, no telemetry, no "anonymized usage data."
A forget button that works. Tell your AI to forget something and it's deleted from disk. Not soft-deleted. Gone.
Auditable in an afternoon. ~2,000 lines of MIT-licensed TypeScript.

Captures where you work

No filing, no tagging, no "save to Memex" button. If you copied it, read it, downloaded it, or screenshotted it — it's in the vault.

📋Clipboardlinks become transcripts & articles
📸ScreenshotsOCR'd on-device
📄DownloadsPDFs parsed on arrival
🌐Browser historySafari · Chrome · Brave · Arc
📝Apple Notessynced automatically

Ask like you'd ask a friend

Vague is fine. Full-text search with real ranking means "that tweet about databases" finds the tweet — in milliseconds, fully offline.

that pdf about python interviews?
📄 100 Python Interview Questions.pdf — downloaded yesterday. Want the section on generators?
what did I copy this morning?
3 items — a GitHub repo on ML interviews, a YouTube link (transcript saved), and a code snippet.

Every AI. One memory.

Claude Desktop and Cursor connect over MCP. ChatGPT and your phone connect over your own tunnel. One vault feeds them all — no lock-in, ever.

ClaudeMCP
CursorMCP
ChatGPTActions
PhoneHTTPS
▲  one local vault · SQLite · ~/.omnicontext  ▲

Rich capture, not bookmarks

Copy a YouTube link and the transcript is indexed before you switch tabs. Articles are archived in full — they survive link rot and paywalls-added-later.

▶
Intro to LLMs — Karpathy
42 min · full transcript indexed
YOUTUBE
🐦
"Postgres is all you need…"
thread text captured
TWEET
🌐
How to Do Great Work
full article archived · 11k words
ARTICLE

Sixty seconds to
a smarter AI

One brew command. Add one block to your Claude Desktop config and ask: "what did I copy today?"

terminal 
# get it
$ brew install arnavbee/memex/memex

# run it forever — starts at login, restarts on crash
$ brew services start memex

# prefer building from source? git clone https://github.com/arnavbee/memex

Questions, answered

Does this send my data anywhere?
No. There is no server to send it to. The optional Cloudflare tunnel exists only so your own phone and your own ChatGPT account can reach your own Mac — behind a bearer token you set.
What about passwords and sensitive stuff?
Every capture path — clipboard, screenshots, downloads, notes, history — runs through the same filter before anything is stored. Copies made while a password manager is frontmost are skipped, as is clipboard content apps mark confidential. Credential-shaped text (AWS/Stripe/OpenAI keys, GitHub PATs, JWTs, connection strings, PASSWORD= lines) is dropped from a clipboard copy outright, and redacted in place inside longer documents so the rest of the file survives. These are heuristics, not a guarantee: a credential rendered as an image and OCR'd, or an unusual key format, can still slip through — the README lists the known limits. Private/incognito browsing is never captured, the vault lives in ~/.omnicontext readable only by you, and anything in it can be deleted by asking your AI to forget it or by running memex purge.
Will it slow down my Mac?
It's a small Node process that polls your clipboard and watches two folders. You will not notice it.
Mac only?
Yes, for now — capture is deeply macOS-native (Vision OCR, AppleScript, launchd). The vault itself is portable SQLite you can take anywhere.
Why not embeddings and vector search?
Personal recall queries are specific — "that PDF about Python interviews" — and BM25 full-text ranking nails those in milliseconds with zero dependencies, zero GPU, zero API calls. If fuzzy semantic recall ever proves necessary, the architecture leaves room for it. Locally, of course.
Is this affiliated with any AI company?
No. It's independent and open source. That's rather the point — the labs are racing to own your memory. You could just keep it.

The best AI is the
one that knows you

Install Memex View source