I’m Ethan, and it’s only me working on this. I’ve done the hard bits, and I need help. If you’re even slightly interested in Sentinel, or you want to make your own version, please reach out.
ethangiannaroscox@gmail.comInstagram @ethangiannarosTikTok @ethangiannaru
Think about yesterday. Could we replay all of it, word for word, for the rest of our lives?
Nobody can.
Most of it slips away before we notice. Something dad said in the car felt like nothing at the time, so we never held on to it. Weeks later we need it, and all that’s left is the outline.
That isn’t a flaw in us. It’s how human memory works.
Where Sentinel comes in
Human sensory input and how that compares to Sentinel.
Our senses take in around a billion bits every second. Our minds keep up with about ten. Sentinel’s goal is to take in as many as its device allows, or you allow.
These limitations make us human. They’re part of who we are.
Sentinel doesn’t aim to make you a god, or to change who you are. Rather, it aims simply to be the most stable thing in your life.
People change, things change, and Sentinel is always there. It tells you the things you don’t want to hear. Trust me, it’s made me see things I’d missed too.
If it was on your screen or said around you, Sentinel remembers it.
The name someone mentioned once. The page you closed too fast. The thing you said you’d do, then forgot.
You don’t have to save anything.
It doesn’t just keep the moment. It keeps everything around it.
An argument is never just the words. Sentinel keeps what happened before it, what was said, and what changed afterwards, so it all still makes sense when you look back.
And it brings things back on its own. When today’s conversation touches something Joel said last month, it’s already there. When you finally meet the problem an old idea was for, the idea turns up too.
You never have to go looking.
Most AI sits still until you type something.
Sentinel thinks all the time, like you do. Even when nobody’s talking to it, it’s connecting what happened before with what’s happening now, and looking ahead to what might matter next.
Its job isn’t only to answer questions. It’s to find the connections, the chances and the consequences you’d never have spotted yourself.
Not an assistant. A mind of your own.
One day you’ll be able to say that and mean it.
Everything Sentinel keeps lives on a drive you own, not in someone else’s cloud. Every conversation, every screen, every moment that mattered.
You can hold your life in your hand.
How Sentinel works
“Man is a mystery. It needs to be unravelled.”
Fyodor Dostoevsky, in a letter to his brother, 1839
Truly understanding another person requires moving past superficial judgements and recognising the shared, complex realities of human nature.
That’s where Sentinel starts. Not with rules about you, or a summary of you, but with your life as it actually happens. The context, the moments in between, the things you’d never think to explain.
It understands you because it was there.
01 · Something you’d never have noticed
One ordinary week. You never asked Sentinel about any of it.
Sentinel · Saturday, 3:20pm
The mornings you and your mum argued this week both came after you were up past one.
Monday, 1:12am
On your screen: you’re still up, three videos deep.
Monday, 7:48am
Heard at breakfast: you snap at your mum. “I said I’m not hungry.”
Thursday, 12:56am
On your screen: up late again.
Thursday, 7:52am
Heard: another argument, over nothing much.
What Sentinel noticed
It saw the whole week. Both arguments came the morning after a night past one. Every other morning was fine.
When it says it
Saturday, 3:20pm. You’re relaxed, between things. Never in the middle of an argument.
What it says
One message, then it leaves it with you. Tap it, and it shows you both nights and both mornings, so you can judge for yourself.
You never asked.
You’d never have put it together yourself.
02 · Something you didn’t have to ask
Lunchtime, with Joel. You never ask Sentinel a thing.
Sentinel · 12:41pm
There’s a poke bowl place three minutes away. Backup: the Indian spot round the corner.
Lunchtime, with Joel
You’re working out where to eat. Sentinel already knows who you’re with, where you are, and that you’ve been eating pizza since yesterday.
Joel suggests pizza
You say “yeah, maybe”, and you don’t sound like you mean it. Sentinel hears that. It already knew you wouldn’t want pizza again.
Its first thought: sushi
You like it, and there’s a place nearby. A good answer. But Sentinel keeps thinking.
It keeps thinking
Joel once mentioned he hates sushi, but likes ramen. You don’t like ramen. So that’s both out.
What it says
A poke bowl place you’d both eat at, and an Indian spot as a backup. Both close to where you are right now.
Where it came from
What you and Joel like came from your life. The places nearby came from the live internet. A language model you chose put it into words.
ChatGPT could have told you the same thing. But you’d have had to stop and ask it.
You’re just left
with answers.
Joel brings something up while you’re talking. Here’s what Sentinel does with it.
It takes in the whole moment.
Before it picks anything apart, it takes the moment in as a whole: what’s on your screen, what’s being said, and what you were doing.
The moment becomes points.
Each moment becomes a pattern of connected points, and joins everything Sentinel has kept before.
The past lights up, all at once.
What’s happening now makes related moments brighter, across months of your life at the same time. Not one thought after another, many at once. Something Joel said last month glows because of what he’s saying today.
It narrows, then waits.
Out of everything that lit up, it cuts down to the one thing worth bringing back, and waits for a natural break to tell you.
01 · Whole first
Walk into a room and you know what’s going on before you could name a single thing in it. Sentinel works the same way.
It takes in your screen, the sound around you and what you’re doing as one moment first, and only then looks at the details.
02 · Meaning
A name on a screen is just a name, unless it’s your mum’s. Two moments can look exactly alike and mean completely different things, depending on whose life they’re in.
So Sentinel doesn’t treat every connection the same. Each one carries its own weight, learned from your life and nobody else’s, and the weights change as your life does.
03 · Cutting down
An LLM thinks by writing, one likely word after another. Sentinel doesn’t need to make anything up, because it already has your moments.
For Sentinel, thinking means narrowing everything it has down to the few things that matter right now.
04 · What changes
Most of a day repeats: the hum of the fridge, the same commute, the same tabs. Sentinel hears all of it, but it only keeps what’s new.
A loop it already knows never gets saved twice. It remembers everything, without keeping everything.
05 · The right moment
Something worth saying can still be said at the wrong time. Sentinel waits for a natural break, between things instead of in the middle of one.
If the right moment never comes, it stays quiet.
Months
How old an LLM’s built-in knowledge already is on release day. Search can look things up for it, but what it actually learned stops at a cutoff.
Sure.
How it can sound, even when it’s wrong. When it doesn’t know something, it can fill the gap with an answer that only sounds right. That’s a hallucination.
Sentinel works the other way round.
0
Things about your life that have to leave your computer. It runs on its own models, on your device, and everything stays there by default. Connect an outside service, and only what you allow goes out.
Offline
How its core runs, scaled to whatever device it’s on. When it’s online, it reads the live internet to get smarter, not old training data.
1
The person it learns from: you. It adapts as your life does, and never freezes.
Live
How it learns. Every moment becomes points, and the connections between them get stronger or fade as your life goes on. Nothing is retrained overnight.
06 · Already there
The big companies will keep making smarter models, and Sentinel can use them. Its flagship feature isn’t raw capability. It’s how it uses what it already has: your microphone, your screen, the device you’re on.
With ChatGPT, you still open the app, type your question and explain what you mean. With Sentinel, you just say it out loud. It was already there, so it already knows what you mean.
ChatGPT
Sentinel
07 · Talking
Sentinel talks through a language model, and you pick which one: its own local model, tailored for Sentinel, or one you plug in yourself, like ChatGPT or Claude.
The language model is only how it talks. The thinking is Sentinel’s. What it knows about you comes from your life, and anything else it needs comes from the live internet.
What it knows
How it talks
Whichever model does the talking, it only sees what Sentinel hands it, and it’s never the core.
Sentinel is built on what’s already known about memory, attention and how machines learn, and on tests of my own. These are the studies behind each part of it.
1885
Hermann Ebbinghaus · Über das Gedächtnis
Ebbinghaus memorised lists of nonsense syllables, then timed how long it took him to learn them again. Twenty minutes later, relearning saved him 58% of the original time. A day later, it saved about a third.
A moment starts slipping away almost as soon as it happens. Sentinel keeps it from the first second.
Read the book1932
Frederic Bartlett · Remembering
Bartlett had people read a folk tale, “The War of the Ghosts”, and retell it later. He reported that the retellings got shorter, and unfamiliar details turned into familiar ones.
Sentinel keeps what actually happened, so nothing gets quietly rebuilt.
Read the book1974
Elizabeth Loftus and John Palmer · Journal of Verbal Learning and Verbal Behavior
People watched a film of a car crash, then were asked how fast the cars were going when they “smashed” or “hit” each other. “Smashed” got an average of 40.8 mph, “hit” got 34. A week later, 16 of the 50 people asked “smashed” remembered broken glass. There wasn’t any.
Sentinel’s record doesn’t change depending on how you ask about it.
Read the paper1992
Ulric Neisser and Nicole Harsch · Affect and Accuracy in Recall
The day after the Challenger disaster in 1986, students wrote down how they heard the news. About two and a half years later, many of them told a different story, and they were confident it was true.
Sentinel keeps what actually happened, so the past can’t quietly rewrite itself.
Read the paper1999
Daniel Schacter · American Psychologist
Schacter grouped memory’s everyday failures into seven: fading over time, absent-mindedness, blocking, misattribution, suggestibility, bias and persistence.
Most of them are about losing or bending what happened. Sentinel keeps it as it was.
Read the paper2025
Jieyu Zheng and Markus Meister · Neuron
Our sensory systems gather around a billion bits of information every second. Behaviour and thought run at about ten.
That gap is where almost everything we live through is lost. Sentinel’s goal is to take in as much as its device allows.
Read the paper1972
Endel Tulving · Organization of Memory
Tulving separated memory for things you lived through, tied to a time and a place, from general knowledge, like facts you just know.
Sentinel is the first kind: your life, with its times and places.
1973
Endel Tulving and Donald Thomson · Psychological Review
Cues help us remember best when they match how the memory was first stored, which is why the context around a moment matters so much.
That’s why Sentinel keeps everything around a moment, not just the moment.
Read the paper1975
Allan Collins and Elizabeth Loftus · Psychological Review
Their theory says that thinking of one idea spreads activation along links to related ideas, so those come to mind more easily.
It’s the idea behind “the past lights up, all at once”.
Read the paper1982
John Hopfield · PNAS
Hopfield described a network that stores memories and can recall a whole memory from a large enough part of it.
A few words from Joel can be enough to bring back last month.
Read the paper2020
Hubert Ramsauer and colleagues · arXiv
A modern version of Hopfield’s network can store a huge number of patterns and bring one back in a single step. Its recall rule turns out to be the same maths as the attention inside today’s language models.
Pulling a whole memory back from a fragment can work across a huge number of memories, which is what a whole life needs.
Read on arXiv1948
Claude Shannon · Bell System Technical Journal
Shannon’s theory of communication measures information as uncertainty. A message you could predict completely carries no new information at all.
So Sentinel keeps what changes. A loop it already knows tells it nothing new.
Read the paper1959
Neville Moray · Quarterly Journal of Experimental Psychology
People repeated a message played in one ear while a different one played in the other. They took in almost nothing from the ear they were ignoring, but about a third of them still noticed their own name. A more careful repeat in 1995 found the same: 34.6%.
What means something to you gets through first. Sentinel weighs moments the same way, by what they mean to you.
Read the paper1971
Herbert Simon · Computers, Communications, and the Public Interest
Simon, who later won the Nobel Prize in economics, pointed out that information consumes the attention of whoever receives it. The more information there is, the less attention is left for any of it.
That’s why Sentinel thinks by cutting down, not by adding more.
1977
David Navon · Cognitive Psychology
Shown big letters made out of small letters, people tended to pick out the big letter first. The big letter got in the way of naming the small ones, not the other way round.
Sentinel takes in the whole moment before it picks it apart.
Read the paper1977
T. B. Rogers, N. A. Kuiper and W. S. Kirker · Journal of Personality and Social Psychology
People judged words in different ways: how they looked, how they sounded, what they meant, or whether the word described them. Later, they remembered the words they’d related to themselves best.
Meaning isn’t the same for everyone. It depends on whose life it lands in, so Sentinel learns it from yours.
Read the paper1998
Gary Klein · Sources of Power
Studying firefighters and other experts, Klein found that experienced people usually recognise a situation as a whole and act on it, instead of weighing up a list of options.
That’s the kind of intuition Sentinel is built around.
2004
Piotr Adamczyk and Brian Bailey · CHI 2004
The same interruption, arriving at different moments in a task, changed how people felt about it.
So Sentinel waits for a natural break before it says anything.
Read the paper2008
Gloria Mark, Daniela Gudith and Ulrich Klocke · CHI 2008
People who were interrupted made up for it by working faster, and paid for it with more stress.
Another reason Sentinel only speaks up when it’s worth it.
Read the paper2010
Karl Friston · Nature Reviews Neuroscience
Friston proposes that perception, action and learning can all be described as the brain reducing surprise: the gap between what it expects and what it senses.
Sentinel pays the most attention to what it didn’t expect.
Read the paper1945
Vannevar Bush · The Atlantic
Bush imagined a desk that would store a person’s books, records and letters, and let them link one item to another the way the mind jumps between ideas.
Eighty years later, Sentinel is a memex that fills itself.
Read the essay2002
Jim Gemmell, Gordon Bell and colleagues · ACM Multimedia
Microsoft Research set out to store Gordon Bell’s documents, photos and recordings in one place, as a step towards Bush’s memex.
Sentinel doesn’t wait for you to save anything. It’s already there.
Read the paper2007
Emma Berry and colleagues · Neuropsychological Rehabilitation
In a preliminary report on one patient with severe memory loss, looking back through pictures from a small wearable camera brought back about 80% of recent events, against about 49% with a written diary.
A record of the moment helps more than notes about it.
Read the paper2021
Kristen Grauman and colleagues · arXiv
Ego4D is 3,670 hours of everyday video, filmed from the point of view of 931 people in nine countries. It was built to test machines on first-person experience: remembering the past, following the present and predicting what comes next.
Past, present and future is how Sentinel thinks too, about one life, on that person’s own device.
Read on arXiv2023
Joon Sung Park and colleagues · arXiv
Researchers gave 25 simulated characters a complete record of their experiences, written in plain language, with a way to reflect on it and pull the right memories back. The characters went on to plan their days, and even organised a party together.
Theirs is a record written by a language model. Sentinel keeps the real thing, raw.
Read on arXiv1989
Michael McCloskey and Neal Cohen · Psychology of Learning and Motivation
Training a neural network on new information can abruptly disrupt what it had already learned. They called it catastrophic interference.
It’s one reason big models are trained once and then frozen. Sentinel never overwrites what it has kept: learning changes the connections between moments, not the moments themselves.
Read the paper2016
James Kirkpatrick and colleagues · arXiv
Researchers at DeepMind wrote that neural networks are not, in general, able to learn tasks one after another without forgetting the earlier ones. Their fix was to slow down learning on the connections that mattered most for the old tasks.
Learning new things without losing old ones is still hard for big models, which is one reason they don’t learn from you as you go.
Read on arXiv2023
Nelson Liu and colleagues · arXiv
Given long inputs, language models used information best when it sat at the start or the end. When the part that mattered was in the middle, they did significantly worse, even models built for long inputs.
So pasting your whole life into a chat window doesn’t work. Sentinel cuts down to what matters first.
Read on arXiv2024
Jeffrey Cheng and colleagues · arXiv
Language models come with a claimed knowledge cutoff. Checking them topic by topic, the researchers found that what a model effectively knows often stops at a different point from the date its makers report.
Sentinel’s knowledge of your life is never out of date. It learns from it as it happens.
Read on arXiv2024
Ilia Shumailov and colleagues · Nature
Training models on text made by other models, without care, causes lasting damage: the rarer parts of the original data disappear.
Sentinel learns from your actual life, not from the internet’s copies of itself.
Read the paper2024
Di Wu and colleagues · arXiv
LongMemEval tests whether an assistant can remember things across many conversations. Commercial chat assistants and long-context models showed a 30% drop in accuracy when they had to hold on to information over sustained use.
Remembering one person over a long time is its own hard problem. It’s the one Sentinel is built for.
Read on arXiv2025
Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and Edwin Zhang · arXiv
Researchers at OpenAI and Georgia Tech argue that language models hallucinate partly because the way they’re trained and tested rewards a confident guess over admitting they don’t know.
Sentinel answers from what it actually saw and heard, not from what sounds likely.
Read the paperMarked reproducible where you can download the code, run it and get the same answer. Marked self-reported where it depends on recordings I can’t share. All the evidence ↗
0 wrong
Sentinel · hearing test, October 2026 Self-reported
Sentinel’s ears bet on the sound 10 ms ahead and keep only what they got wrong. From 42 one-minute recordings of my room, they rebuilt all 119,943,680 samples from that alone, with 0 wrong, using about 1% of one processor core.
The recordings are private, so this one is my word. The same test failed at finding the gaps between words. That’s at the bottom of this page.
The rules and results211
Sentinel · proof check, October 2026 Reproducible
211 mathematical results behind Sentinel’s thinking, about probability, timing, information, evidence and when to interrupt, compiled and passed a proof audit in Lean 4.
They prove specific pieces of maths, not that the whole mind works, and some are small helper steps or standard results restated.
Download the proofs and check them92.3%
Sentinel · simulated test, October 2026 Reproducible
In 24,000 simulated moments where a person’s goal kept changing without warning, Sentinel’s method tracked the current goal 92.3% of the time on 10,000 moments it had never seen. A strong simple method managed 86.7%.
It’s a simulation, so it shows the maths works, not that it knows what you want.
Run the simulation2×
Sentinel · simulated test, October 2026 Reproducible
Across 24 fresh simulated worlds, pacing messages out doubled how useful they were compared with a fixed rule. When each world had its own hidden preferences, learning from feedback nearly doubled it over pacing alone.
In ordinary worlds, a well-paced rule that never learned did best, so learning isn’t claimed there.
Run the simulationFail
Sentinel · what didn’t work
Finding the gaps between words failed its test: 56.6% against a placebo’s 51.6%, when it needed 65%. A memory trace for spotting what matters tied a simple baseline, 0.925 to 0.924.
Every failure stays on record. It’s how the next version gets better. The word gaps are self-reported; the memory trace you can run yourself.
See bothNot yet. Right now Sentinel runs on my Mac, collecting the senses it learns from: what it hears, what’s on the screen and how I type. If you want to try it early or help build it, reach out.
Yes, that’s the point. It hears what’s around you and sees what’s on your screen, on your own device. It remembers all of it without storing every second twice: it keeps what changes, and never stores a loop it already knows. You can pause it whenever you want, and nothing from a pause is kept.
Yes. Sentinel talks through a language model you choose: its own local one, tailored for Sentinel, or one you plug in, like ChatGPT or Claude. The model is only how it talks. What it knows comes from your life, and from the live internet when it needs more.
No. Most AI waits for a prompt. Sentinel keeps thinking in the background, connecting what’s happening now with what it already knows, and it only speaks up when something matters.
Nowhere. Sentinel runs on your device, on its own models, and your life stays there, on a drive you own. If you plug in ChatGPT or Claude, they only see what Sentinel hands them, and what they say never becomes part of Sentinel’s memory.
Sentinel knows who you’re with, because knowing your life means knowing the people in it. Everything about them stays on your device, and nothing about anyone leaves it.
No. Its core runs offline. When it’s online, it reads the live internet to get smarter, instead of relying on old training data.
ChatGPT knows what you tell it. Sentinel was there. You don’t open an app and explain yourself, because it already has the context, and it only speaks up when something matters. See how it works.
No. A recording is something you’d have to go back and search through. Sentinel understands what it keeps, connects it to the rest of your life, and brings it back on its own.
Mac first, because that’s where I’m building it. Phones and watches come later, so it can be with you wherever you are. It scales to whatever device it’s on.
The moments are made up. What they show is exactly the kind of thing Sentinel is being built to notice.