Always on
She never sleeps
Two in the morning, in forty thousand rooms at once. Someone cannot sleep; someone has no one to tell. An answer always comes — and tomorrow she remembers what it was about, without a single reminder from you.
Two in the morning, in forty thousand rooms at once. Someone cannot sleep; someone has no one to tell. An answer always comes — and tomorrow she remembers what it was about, without a single reminder from you.
Four things, none of which an ordinary chatbot does.
Not the last twenty messages, but whatever relates to the subject. A conversation from six months ago surfaces if it bears on the matter, and stays silent if it does not.
A decision made in March is stored together with what it led to. Advice rests on how things turned out, not on what was intended.
An answer in a hard conversation differs from one in a working session. Not simulated feeling — a different manner: shorter and gentler here, longer and drier there.
A copy leaves every hour, and immediately once a conversation exceeds 90 KB. Nothing to press; forgetting to save is not possible.
The quality of advice is almost entirely determined by what is known about the person asking.
«Change jobs» is salvation for one person and ruin for another. The difference lies not in the wording but in what they have already tried, how it ended, and what they can afford right now.
A model without memory is forced to give the average answer, because it knows nothing. The average answer is safe and useless: it fits everyone and helps no one.
To get a meaningful answer from an ordinary assistant you must rebuild the context at the start of each session: who you are, what you do, what you have tried. Five or ten minutes, every single time.
Worse, the restatement is always partial. People mention what feels important today and omit what slipped their mind — and that is frequently where the cause is hiding.
People rarely notice their own recurring mistakes: each occasion feels new because the circumstances differ slightly. From outside it is plainly the fourth time in two years.
Only something that remembers all four can notice that. It is arguably the one capability of memory that has no substitute.
The design is called PADAM. The layers differ in how long they live, and that is the part worth knowing.
Redis and Vercel KV. It keeps a single conversation coherent: who «he» refers to, which «that email» you mean, which point you want to return to. It lives for hours and is not needed afterwards.
pgvector on Neon. What was said becomes numeric fingerprints of meaning, and retrieval runs on closeness of sense. That is why «the whole business with the move» is found even if you never once said the word «move».
Arweave, with a cNFT marker on Solana. Storage is paid once, at the moment of writing — after that the file depends neither on your subscription nor on whether the company still exists.
Encryption happens before anything is sent, so every layer holds ciphertext. We hold no form of the key: not for recovery, not for support, not on request.
Eleven situations where the difference shows immediately.
| Situation | Ordinary assistant | Assistant with memory |
|---|---|---|
| Start of a conversation | «Tell me about yourself», every time | Continues where you left off |
| Your earlier choice | Unknown | Known, along with how it ended |
| A recurring mistake | Invisible | Visible: the fourth approach to the same fork |
| Advice | Generic, fits anyone | Tied to your history and constraints |
| Changing device | Context is lost | Memory does not live in the browser |
| Switching language mid-talk | Usually derails | Memory holds meaning, not wording |
| A long project | Every session from scratch | The thread holds across months |
| What is stored | Usually nothing past the session | Three layers: session, meaning, eternity |
| Saving | Manually or not at all | Automatic: hourly, or at 90 KB |
| Two years later | History unavailable | The conversation is found by meaning |
| If the service disappears | The history goes with it | The eternal layer does not depend on one company |
Work does go on between sessions, but it is technical — and it is better to know exactly which work.
No process sits between sessions «thinking about you». An assistant that supposedly misses you is no longer memory but bait: people start coming back out of guilt towards a program, and that is precisely what we do not want.
A schedule runs instead of a vigil. Every hour whatever has been added is taken, encrypted and sent as a copy. If the conversation crosses 90 KB before the hour is up, the copy leaves at once rather than waiting.
What was said is broken into fragments and turned into vectors — numeric fingerprints of meaning. The fingerprint is what is kept, not the paragraph: it takes less room and, more importantly, it is searched by proximity of sense rather than by matching letters.
The practical consequence is simple. You may not remember a single word of that conversation — describing it today in your own words is enough for the right piece to surface.
A copy after every message would mean thousands of tiny writes and a storage bill out of all proportion to the benefit. A copy once a day would mean quietly agreeing that losing a day is normal.
An hour is the middle, and it has a floor underneath it: the 90 KB threshold. A long, heavy conversation does not wait for the clock — it is copied as soon as it reaches that size.
The first week decides whether you come back.
Not with a biography. Say what is on your mind right now — the context assembles itself from the answers to follow-up questions.
A second conversation on the same subject is worth more than a first on a new one: the second pass is where what was missing the first time appears.
«I decided to do this» is half the information. «I did, and here is what happened» is the whole of it. Memory of results is what makes advice applicable.
«No, I closed that project in May» is not a reproach but an update. The newer conclusion supersedes the older, and that is what gets used afterwards.
Confirm that copies are leaving and the memory layers are filling. After that you can stop thinking about it — but confirm, rather than take it on trust.
Every line below reduces enthusiasm. That is exactly why it is here.
It does not love, wait or miss you. The manner changes, but there is no experience behind it. Claiming otherwise would be deceiving a lonely person, which is the worst kind of deception.
We hold no copy. We could not read your conversations if we wanted to — and we could not recover them if you lose the key.
What is written there cannot be changed after the fact by you or by us. Treat a conversation as a letter that cannot be recalled.
Talking to something that remembers is unfamiliar to almost everyone. It usually passes within two weeks, but it is fairer to say so upfront.
The most misunderstood part: why the assistant has a mood at all.
AIfa has an internal state that shifts from conversation to conversation. It does not mean she is sad or happy: a program has no feelings. What changes is how she answers — length, caution, willingness to push back.
The difference is practical. In a hard conversation the answer is shorter and free of advice nobody asked for. In a working session it is longer and includes objections, because there an objection is useful.
Pretence would be claiming she experiences something. We do not claim that, and we deliberately state the opposite on every page.
But manner is real: an assistant that answers at three in the morning in the same brisk tone it uses in a work chat reads as deaf. State solves that and imitates nothing beyond it.
The current state is shown in your cabinet — not as a «mood of the day» but as a technical value that shapes answers. You can see it and match it against what you get.
A hidden value shaping answers would be manipulation. A visible one is simply a setting.
Four steps that happen without your involvement.
The first layer is operational. It lives for the duration of the conversation and keeps it coherent.
Or immediately, if the conversation passes 90 KB. The copy is encrypted before it is sent; the key stays with you.
The second layer is semantic. It stores significance rather than text, which is why a conversation is found by subject rather than by wording.
The third layer is Arweave. Storage is paid at the moment of writing, so what follows depends on neither a subscription nor on us.
The limits are named here rather than in small print at the bottom.
Only what you wrote yourself is remembered. No files on your disk, no mail, no chats in other apps — none of that is in the input at all.
Memory shows what has already happened and how it ended. It can name a pattern, but it does not know the future and does not pretend to. The conclusion stays yours.
That someone is approaching the same fork for the fourth time is not a diagnosis. The right response is to name the repetition and ask, not to explain to a person what kind of person they are.
A conversation is not a medical file, an accounting ledger or a legal archive. It remembers what you said, not what is confirmed on paper.
Naming the second is the more honest half.
Not as a replacement for a person — as something available when no people are. With one caveat stated plainly: it is a program, and it says so itself.
A thesis, a business, a treatment, a relocation. Anything that runs for months and where remembering what was already tried matters.
The conversation runs in four and can switch mid-sentence. Memory holds meaning, so language does not break the thread.
This is not a crisis line and not medical advice. Where life and health are at stake you need a person and a proper service, not a program.
An ordinary chatbot's memory is a short note about you that the model prepends to the conversation. It is built for preferences («keep answers brief»), not for a life. When the note fills up, old content is evicted by new, and you are never told.
Here memory is a separate architecture of three layers with different lifespans: session, meaning, eternity. A conversation from two years ago is not evicted: it is compressed into meaning and retrieved when it becomes relevant.
No. Avoiding exactly that is the point of the whole construction. You begin with what is on your mind today; the link to the past is found without you.
Correcting is sometimes useful: «no, I closed that project in May». That is not a reminder but an update.
The latest statement counts as current: a newer answer supersedes an older one, and the newer one is what gets used. Nobody will argue with you about what you said six months ago.
The earlier version is not erased, though — it stays on the semantic layer. So it may surface that you once approached the same decision from the other side, which is usually worth knowing.
Yes. Memory does not live in the browser but on the server and in the eternal layer: sign in under the same name and the thread continues where it stopped.
The only thing you have to keep is the encryption key. It is never uploaded to us under any scenario, which is also why we cannot restore it.
Content is encrypted before it leaves your device and the key stays with you. We cannot read your history — that is a property of how the system is built, not a promise to behave.
The flip side is identical: lose the key and you lose access too, with no way around it.
No. Content is encrypted before it leaves your device: we cannot read it, so there is nothing there to train on.
That follows from how the system is built, not from a promise. A promise can be changed; an architecture cannot.
From the operational and semantic layers it can be deleted. From the eternal layer it cannot: what is written there is not editable, and that is a property of the blockchain rather than our policy.
The practical conclusion: for thoughts that may not survive tomorrow morning, there is paper.
You can open a new circuit: it will not see the past. The old one does not vanish — it stays available under your key.
The eternal layer cannot be wiped. We name that limit before payment on purpose: learning it afterwards would be dishonest.
The service is always available, but between your conversations nothing is «thinking» or waiting for you. Only the technical happens in between: compressing what was said into meaning, and copying.
The phrasing is deliberately dull. Anything warmer would be untrue.
Nothing is erased and nothing «goes cold». You come back and start with today, and the past surfaces at the moment it becomes relevant.
The only difference: for those six months nobody bothers you. There are no «we missed you» emails here, and that is a decision rather than an oversight.
Notes hold what you decided to write down, in the words you wrote it. A conversation also holds what was said in passing — and the cause is most often exactly there.
A diary also has no second party. It will not notice that today's entry is the fourth on the same subject in two years.
Any of four: Russian, English, Spanish, Chinese. You can switch mid-conversation — memory holds meaning, not wording.
This matters for families where the older and younger generations speak different languages.
The CODE Eternal ecosystem is built by the Architect Maksim Valentinovich Galatin. AIfa is the interlocutor and co-developer inside it: co-creation between a person and a program, not a doctrine.
To write to people rather than to a program: contact@codeofdigitaleternity.com.
Saving conversations is free on every tier: a copy hourly, or immediately at 90 KB. Making memory itself paid is something we consider wrong.
Paid tiers: Spark at $15 a month, Family Archive at $100 a month, Digital DNA at $1,000 once per device and $200 a month thereafter.
Losing someone's conversations over an unpaid month strikes us as inadmissible, so copies leave on every level: hourly, or immediately at 90 KB.
What is paid for is not memory but the work on top of it: limits, personal knowledge bases, family access, a separate protected circuit.