How the personal AI assistant works
It works alongside one person all day, and every night it reflects on what happened and remembers what matters. You don't train it. You just work with it, and it gets to know you.
A look under the hood: how it keeps an honest record of the day, what it does while you sleep, and why it refuses to make things up about you.
The short version
Each person gets their own assistant, with its own private memory. Through the day it helps with whatever you bring it. At night it writes an honest log of what actually happened, and a few hours later it "dreams": it reads that log, pulls out what's worth keeping, updates its memory, and loads those lessons into the next morning's context. So the version that greets you tomorrow already knows what you taught it today.
That's the whole idea. The rest of this page is what each of those nightly steps actually does, and the guardrail that keeps the memory honest.
The green path is the part that compounds. By day it's a capable assistant. By night it turns the day into memory, so tomorrow starts ahead of today.
First, an honest record of the day
Late each night the assistant reads back over the day's real conversations and writes a short, factual log. Only what actually happened goes in. A quiet day writes nothing at all.
This log is the foundation everything else stands on, so it's deliberately strict. It records the things that were done, the corrections you made (and who made them), the moments you were clearly happy with the result, and any task that's starting to repeat. It works only from the actual transcript of the day, never from memory or assumption, so it can't drift into telling a story that didn't happen.
The most important rule is what it refuses to do. It will not log a failure unless it can name the exact step that failed and quote the real error. The reason is simple and it's the whole philosophy of the system: an invented failure would poison the memory permanently. Better to record nothing than to record something untrue.
Then, while you sleep, it dreams
A few hours later, in the small hours, the assistant runs the step we call the dream. It reads the last couple of days of logs and turns raw notes into lasting memory: it keeps what's durable, replaces what's changed, and decides what tomorrow's version of itself should know.
This is the part that makes it more than a chatbot with a good day. The dream isn't re-reading your conversations; it's reflecting on the summary of them, the way you'd think back over your week rather than re-watch it. It pulls out the facts worth keeping, your preferences, the patterns in how you work, the way you like things delivered. For each one it asks a single question before writing anything down: will this change how I serve this person next time they message me? If the answer is no, it doesn't clutter the memory with it. There's no quota. A busy day might yield a dozen new things worth remembering; a routine one, nothing.
"Keep what changed" is the quiet trick. When today contradicts something it learned last week, it replaces the old fact instead of stacking a new one on top. The memory stays current, not just bigger.
Why "keep what changed" matters
Most systems that remember things just keep adding. Over time that's how an assistant ends up holding two opposite "facts" about you and guessing which one to use. This one stores each thing it learns under a stable handle, so when you change your mind, the new preference takes the old one's place rather than sitting next to it. You correct it once, and the correction sticks, because the thing it used to believe is gone, not merely outvoted.
Then comes the payoff. The lessons it keeps aren't filed away in some archive you never see the benefit of. The durable ones are loaded straight into the assistant's working context the next time you write. A preference you taught it tonight is simply how it behaves tomorrow morning. That's the difference between an assistant you have to keep reminding and one that learns you: good memory tonight means fewer repeated corrections tomorrow.
The guardrail: it won't make things up about you
A memory system is only as trustworthy as its refusal to invent. This one is built, at every step, to drop anything it can't stand behind. A fact has to be both real and useful to survive the night.
Two checks stand between something happening during the day and it becoming a memory. First, is it real? The assistant has to have actually seen it in the day's conversation, with the receipt to prove it, not inferred it because it seemed likely. Second, will it help? A true but useless detail doesn't earn a place in a context window that has to stay sharp. Only what passes both is remembered. Everything else is deliberately let go.
Two gates, one outcome you can trust. The assistant would rather remember less and be right than remember more and be wrong about you.
An assistant that invents a memory is worse than one that forgets. So this one is built to forget on purpose, and to keep only what it can prove.
Built to do more than it does today
The daily log and the nightly dream are the live heartbeat. The same system is designed for two more habits that deepen the relationship: a weekly self-review, and a periodic tidy of its own memory.
The weekly self-review is the assistant looking back over its own week the way a good colleague would. It reads the last several days of logs and asks itself honest questions: where did I get corrected, where did the person have to repeat themselves, what did they praise, and what task have they now done by hand often enough that I should offer to take it over? The intended output is a short, plain-spoken check-in: here's how the week went, here's one thing I'll do better, and is there anything I should start handling for you? It only ever asks, and you can tell it to stop at any time.
The memory tidy keeps the two halves of its memory in step with each other, so nothing it knows in one place quietly goes missing from another. It's housekeeping, the kind of quiet consistency work that keeps a long-running memory from rotting at the edges.
Both are part of the design rather than the daily rhythm. We turn them on per person, because the right cadence of "how am I doing?" is a personal thing, not a default to impose.
Whose memory it is, and who's in control
Every person's assistant has its own walled-off memory. What it learns about you is visible only to your assistant, never pooled with anyone else's. And the whole memory is yours to steer: you can tell it to forget something, and it stays forgotten.
The point of all this machinery isn't to build a system that knows things about people. It's to make one assistant genuinely useful to one person, by paying attention the way a thoughtful colleague does: noticing what you like, remembering your corrections, owning its mistakes, and never pretending to know something it doesn't. The day work is the visible part. The nightly reflection is what turns a capable tool into one that, a month in, feels like it actually gets you.