Inspiration

"Wait, what time did we say?" We all scroll up through weeks of WhatsApp to find one promise, one address, one deadline. Our chats already hold the answers, we just can't reach them while typing. We wanted a memory that sits where you type and stays private.

What it does

Recall is a Chrome extension with a private memory of your chats:

Suggestions in your style: grey text that finishes your message the way you write; Tab accepts.
Recall: when a friend asks "so when are we meeting?", a card shows what you said before, with the date and the original message ("You said: tuesday 5 pm works for me, 12 Oct"). If the answer isn't in your chats, it stays quiet instead of guessing.
Reminders: type "ok I'll send the slides tomorrow" and Recall offers a reminder (Chrome notification + Google Calendar link).
Privacy panel: see and delete everything Recall remembers. OTPs, passwords, PINs, card and bank numbers, Gov IDs are hidden on your device before anything is sent.
Clicky: an on-screen guide for any website. Ask "how do I share this file as view only?" and a cursor points at each step. Modes: Guide me, Do it with my OK, or Autopilot, which stops at checkpoints and never presses Send, Pay or Delete for you. Clicky can also take questions by voice (Whisper runs in the browser), fill a form from your chats ("fill my address from my chats", showing which chat it came from), carry selected text to another tab, and show a plan checklist.

How we built it

Chrome extension (Manifest V3, plain JavaScript). Chats are imported from WhatsApp exports; memory and embeddings live in chrome.storage.local, never on a server.
Stateless FastAPI backend on Render that calls NVIDIA Nemotron on Nebius Token Factory: Nemotron 3 Nano for live suggestions (speed), Nemotron 3 Super for recall, reminders, Clicky's guidance, plans and form filling (accuracy). Qwen3-Embedding-8B on Token Factory for memory search.
Retrieval: embed the question, find the closest past messages in the browser, add the messages right after each match (answers usually follow questions), and send only those ~30 messages to Nemotron Super.
Whisper tiny runs inside Chrome via transformers.js, so voice never leaves the computer.
A website with a live demo: the extension's own chat page running in the browser on a made-up chat, against the same backend.

Challenges we ran into

Empty answers: Nemotron 3's reasoning mode spent the whole token budget thinking. Turning reasoning off for these short, structured tasks fixed it.
Recall picking the wrong message, e.g. quoting the friend's question instead of your answer. We added the follow-up messages to the search results, set temperature 0, and wrote rules for who said what.
Suggestions that were just filler ("bro lol") or invented numbers. We filter both out and measured the change.
Safety in Autopilot: deciding where an agent must stop. It always pauses before Send/Pay/Delete/Submit and before Save/Done.
The plan checklist made guidance worse when we fed it back into the guide prompt, so we kept the guide unchanged and tick the checklist off in the extension instead.

Accomplishments that we're proud of

Privacy by design: memory stays in the browser, secrets are hidden before anything is sent, the server stores nothing.
Recall shows its source (date + original message) and says nothing rather than guessing.
We built a 40-case evaluation with a Nemotron Super judge; filtering filler and made-up numbers cut wrong suggestions from about 40% to about 28%.
Clicky works on any website and turns Recall's memory into actions (filling forms from your chats) with your OK at each checkpoint.

What we learned

For short structured outputs, prompt rules + temperature 0 + JSON-only replies matter more than model size, and reasoning mode should be off.
Retrieval quality (what you send the model) decides most answers.
Measuring with an eval set beats judging by feel.

What's next for Recall

Run inside WhatsApp Web and Gmail directly (today the chat features run in Recall's own chat page; Clicky already works on any website).
Chrome Web Store release.
More languages for voice and Hinglish-aware suggestions.