We write what we learn.
Product, tech, and design — how a voice assistant with neural memory is built.
XNeuronal is available on Android
A personal assistant with a memory, built on the way a human memory ranks, connects and brings things back. What it does, what it does not do, and why there is no tab and no folder to create.
You asked for it: reading your phone calendar
The briefing said “no appointments”; the phone calendar had some. How XNeuronal learned to read that calendar, and to write what you dictate back into it.
A memory that works things out is almost here
Our note taking tools keep words, not memories. They know nothing about who, when or where, and they never come back on their own at the moment you need them. Here is what we built instead, and how to hear about it the day it ships.
A memory that does not make up who people are
Until 24 May, a person you mentioned without saying who they were ended up filed somewhere, often among your friends. Since then, the app knows how to say "I don't know yet": one area of life for people who were only named, another for acquaintances, an automatic link between things said in the same breath, and links that have a name. And, the next morning, pitch-black dots on the map.
What stays when you delete your account
On 24 May, XNeuronal got a "Delete my account" button that really erases everything on the server, and the big clear-out in the settings stopped leaving reminders and transcripts behind. The story of one morning spent on a family question: what you keep, what you erase, and what belongs to whom.
A memory that does not repeat itself
You tell the same thing three times in three weeks, in different words. Until 22 May, the app kept three separate memories of it, and the mechanism meant to recognise them had never worked once. What was repaired, how a memory is recognised under another wording, and the two gestures that let you tidy up yourself.
A constellation that grows as you talk
Until 22 May, the middle of the home screen showed the same orb to everyone, a demonstration that never changed, whatever you told it. Since then, it is your memory that draws itself there, point by point, from the very first memory. How the map is animated, why the twenty-memory threshold did not survive the day, and two other fixes from the same day: formatting inside cards, and the question "what have I got next week?".
What the app chooses not to remember
Tomorrow's weather, the traffic, last night's score: if everything you ask became a memory, your memory would be a drawer full of old receipts. How the app learned to show without keeping, while still leaving the information within reach for as long as the conversation lasts.
When the assistant's instructions ran to a thousand lines
A thousand lines of rules, counter-examples and special cases, reread before every reply, and an assistant that lost the thread two sentences later. On 21 May everything was rewritten as a short specification that says what is true rather than what to do. What that changes when you talk to it.
What were we talking about last time?
The long-term memory held, but the thread of the conversation was lost every time the screen went dark or another app took over. How, in one afternoon, the assistant stopped looking amnesiac when its memories were intact.
When it isn't you talking
You hand the phone to your wife, your daughter, a friend, and everything they said used to land in your memory as if you had said it. How the app learned to tell who is talking, who it credits with what it keeps, and why the owner never changes.
The conversation that survives dropouts
You are talking to the app on the train, the network drops for ten seconds, the link comes back on its own. Before 19 May, it came back having forgotten what you were talking about. How the thread of a conversation stopped belonging to the cable and started belonging to the person, and why a passing error no longer wipes it.
When one sentence holds three memories
"My wife's name is Claire, my daughter is eleven, my son eighteen": an ordinary sentence, three things to remember, and nothing remembered. How, in one evening, the memory learned to count the facts in a sentence, to stay quiet when it retains, and to light up a whole neighbourhood from a single cue.
A briefing you can touch
Every morning the app tells you “you have three events today”. Then what? How that summary became a door to the first appointment, how the calendar's month view stopped running away when you tap a day, and why any text that names a thing should lead to it.
The links were there, nobody followed them
A memory dutifully wrote down the links between its own recollections, then never read them back when it came time to remember. How one hop through the graph, a threshold of three, and a map pinned to the top changed what it can find.
The card you can fix by hand
A dictated card is never quite right. How it became editable with a fingertip, from the misheard quantity to the checkbox Android refused to draw.
A picture for every card, and why it got it wrong
A recipe deserves a photo, not a drawing. How every card got its picture in one afternoon, and why it first showed everything but the dish itself.
When the assistant says "OK, I'm looking" before answering
Five seconds of silence look like a crash. How, in one day, the screen and the voice learned to say "I'm listening", "I'm thinking" and "I'm looking".
Filling slots instead of writing text
A note that comes back as a wall of text forces you to reread it all. How we reversed the relationship: the screen declares a template, memory fills the slots.
A list without a tab
You had to be able to look at your memos and your tasks. You could not add a tab. How a sideways swipe settled the contradiction, and why cards stopped arriving empty.
The account that keeps you from losing your memory
You can use XNeuronal without an account, and that is deliberate. But a memory tied to a single phone disappears with it. The story of a piece of work on email confirmation, forgotten passwords, and how to invite without forcing.
Your settings are part of your memory
A choice that disappears on reinstall was never remembered, just stored somewhere. Why preferences left the phone and moved into the memory itself.
Voice-powered web search
An AI that answers a question without checking its facts will confidently invent. Here's why we've integrated real web search, with cited sources, directly into your conversations.
One Dock, Two Modes
Switching from voice to text shouldn't be a guessing game. Here's how a single small dock, with two nodes connected by a thread of light, replaced a shape-shifting button and a swipe you had to learn.
The Real Push-to-Talk
A button that cuts you off isn't real push-to-talk. Here's why we moved away from phone-based speech recognition to redefine what "holding the button" truly means.
A Tiered Memory
A memory that keeps everything eventually drowns in its own noise. Here's how we introduced tiers (and an inbox) so XNeuronal knows what to prioritize for you.
Real-time, no refreshing
A conversation isn't a series of forms. Here's why XNeuronal relies on an open channel, not isolated requests.
Building a Voice: Speaking, Understanding, Responding
Behind a simple conversation, there are three layers of processing and a commitment to simplicity. Here's how we built the voice pipeline.
Where did the idea of a memory that organizes for you come from?
Capturing information has never been the problem. Finding it again, that's the real challenge. A look back at XNeuronal's initial intuition.
Why We Removed Tabs
Productivity apps pile on screens. XNeuronal did the opposite. Here's why a single page changes everything.
How Neurons Work
Everything you say becomes a node connected to others. Let's explore our vector memory architecture.
Designing an Invisible Interface
The best UI is the one you never use. A look back at 14 iterations of the central orb.