Private beta / Bring your own model
It remembers what matters, lets go of what doesn't, and you can correct it.
AI chat that keeps notes on your work — typed, fading over time, and yours to edit.
What it keeps
A memory is a sentence you can read, question, and change.
Not a hidden embedding. Each one carries what it says, where it came from, and whether you have corrected it — so you can disagree with it.
- Decision
Ship the private beta to five friends before building the mobile app.
from a chat about scope, 3 weeks ago
- Friction
Import kept stalling on large exports — fixed by paging the extraction window.
from a debugging session, 8 days ago
- Corrected by you
Prefers Postgres over SQLite for anything with row-level security.
you edited this — the old wording no longer applies
Illustrative examples, not captured from a real account.
Trust
Control, not magic
Your model connection
Connect OpenAI, Anthropic, Google, or an OpenAI-compatible endpoint, billed by that provider at their rates. A limited number of invited testers instead run on a capped, Refinery-funded allowance.
Visible context
Citations, memory, files, and context receipts stay visible so you can inspect what an answer used.
Protected by design
Provider keys are encrypted and app data is protected by Supabase row-level security. Refinery does not claim certifications.
Private beta
Bring memory into the next conversation.
Request access, then connect your own supported model provider for beta use.
FAQ
Do I need an API key?
For most beta access, yes: connect your provider key before full chat and source indexing work. Memory features specifically need an OpenAI key today. A limited number of invited testers get a capped, Refinery-funded allowance instead and can skip this until it runs out.
What does BYOK mean?
Bring your own key: you connect your own OpenAI, Anthropic, or Google account and that provider bills you directly. That's the default for beta. A small number of invited testers instead run on a capped tester allowance funded by Refinery, up to a fixed cap, not unlimited or hidden.
Is my data private?
Provider keys are encrypted, app data is protected by Supabase RLS, and you can already inspect, edit, or delete individual memories and see the sources behind any answer from the context panel. Refinery does not claim certifications.
Does it work on mobile?
The landing page is mobile-responsive. The product is desktop-first during beta; mobile app workflows come later.
How much does beta cost?
Default beta access runs on your own model key, billed by that provider at their normal rates. A limited number of invited testers instead get a capped, Refinery-funded allowance at no cost to them. Hosted usage and paid plans are planned separately.
What models are supported?
Chat runs on OpenAI, Anthropic, Google, or a custom OpenAI-compatible endpoint. Two things about memory are worth knowing. Extracting memories from a conversation needs an OpenAI key today — without one, chat works but nothing new is remembered. Recall differs too: with an OpenAI key it is semantic, finding a memory that means the same thing in different words; without one it is keyword-based, finding memories that share your wording. Refinery tells you which mode you are in rather than hiding it. Capture, typing, editing and deletion are identical either way.
Can I delete or export data?
Existing export/delete surfaces are present but still need a full data-rights matrix. Public copy stays modest until that work is complete.
Is canvas available now?
Canvas exists as a project work surface but is not the landing-page promise.