A Working
Document on ARIA.
Most AI has a context window. ARIA doesn't. A reasoning layer built around the long arc of your thinking — what you have written, decided, and changed your mind about.
ARIA is a reasoning layer that holds the long arc of your thinking. Most AI runs on a context window — a fixed amount of recent conversation it can hold in mind. When the window closes, the model forgets. ARIA was built around the absence of one. There is no point at which it forgets the beginning.
This document is what that feels like, after a year of using it daily.
01 — On the problemThe interesting problem is not memory.
Most AI tools reset between conversations. You re-explain. You re-paste context. The model has no memory of what you decided last week, what you abandoned last month, or what the constraint actually was. The interesting problem — the one almost nobody is solving — is continuity.
A person's thinking, accumulated over years, is the most valuable context any tool could have access to. The decisions. The reasons behind them. The notes you wrote at 2am that were sharper than the ones you wrote in the morning. The thing you almost said in a meeting last March and then didn't.
Search gets you halfway. Long context windows get you the next quarter — but they ceiling out the same way. There is always a beginning the model can no longer see. ARIA does not have that ceiling.
"We have built impressive tools for short interactions. We have built almost nothing for long ones." — field note, 2025-12-04
02 — What it feels likeThree things, on a good day.
The clearest description is the experience. Most of my use is one of these three:
- I ask about something I already decided. It tells me when, what, and the reason I gave at the time. If I have contradicted myself since, it surfaces both versions and asks which I meant. I do not have to remember where I wrote it down.
- I wake up to a brief. Three to five things from the last day worth my attention — not because they are new, but because I had marked them as load-bearing without realising. The reply I owe. The decision that has been rotting open for a week. Two notes that look like the same idea.
- I start writing about something old. It pulls in the earlier writing — including the version where I thought differently. I either change my mind or sharpen it. Either way, I do not repeat the same thinking.
Most other behaviour is downstream of these three. The interface is small on purpose.
Three things worth noticing.
- 01 The reply to M.K. is still in your drafts. Six re-opens. Most recent: three days ago, two minutes.
- 02 You almost-decided on the doc layout twice and reversed once. The current draft keeps the marginalia. Worth committing or revisiting today.
- 03 Your reading on retrieval overlaps with last fall's thinking. You flagged this in November; nothing has been written down since.
03 — Live demoReplays of a real session.
Below are six queries I have run during normal use these past few weeks, replayed exactly as ARIA returned them. The internal mechanics are not shown — the answer is the only part that is meant to convince. Read a few. The specificity is the point.
What is striking is not the speed. It is the specificity — the answer references things I had genuinely forgotten and would not have found by searching, because I would not have known what to search for.
04 — What it is notNegative space.
Defining a system by what it is not is usually clearer than the alternative.
- Not a chatbotSessions are not the unit of interaction. There is no "new chat" button. After a few weeks of use, you stop thinking of it as a chatbot at all — it feels like talking to someone who has been quietly watching the way you think for a while.
- Not a productivity appNo inbox, no kanban, no task list, no streak counter. ARIA does not want your attention. It wants to spend yours wisely on your behalf.
- Not training dataWhatever ARIA sees stays where it lives. The platform never sees your inputs in plaintext, and reasoning happens in places configured to remember nothing across calls. Your context is yours.
- Built around youEach ARIA learns one person — yours learns you. The longer it watches, the sharper its picture of how you think becomes. That picture is yours, not a dataset.
- Still movingIt is alpha, and it is improving fast. Each version has been sharper than the one before, and the shape is still settling.
05 — Where this standsIn motion.
ARIA is in active development. Everything described above is what it does today — running, daily, for one user. More is coming through 2026: new behaviours, the app you would actually use, public access. The next version of this document will be the place to find out what is open and how to try it.
Until then, if you want to talk about the ideas — the design, the philosophy, where this is headed — write any time. I read everything.
If anything here resonated, the rest of the conversation happens in email.
Email Anshul → ar.dev@anshulraman.com