How kruel.ai Thinks

Abstract. kruel.ai is a single persistent mind that lives on hardware its owner controls and reaches them through many bodies: a desktop, a phone, a browser, a message thread, a headset, a physical robot. Its defining property is not that it remembers, which many systems now do in some form, but what kind of remembering it is. Everything the mind experiences, whether spoken, seen, read, or typed, becomes part of one life, held with when it happened and why it mattered, and recalled under a rule written for people whose own memory is failing: a weak but true memory must always remain reachable. On top of that memory sit habits of thought that a person would recognise: a fast hunch and a slow deliberation, sleeping on things, learning from outcomes, questioning its own beliefs, and refusing to say it did something it did not do. This paper describes those habits in plain language, places the work in the lineage of thinking machines, and states honestly which ideas are ours, which are shared with the field, and what is still open.

01Why we explain it this way

Most descriptions of an AI system are descriptions of plumbing. That is useful to engineers and useless to everyone else, and it hides the only question that matters to a person who might live with one of these: what is it like to be known by it?

So this paper uses the vocabulary of minds. Where kruel.ai does something a human mind also does, we say so and use the human word. Where it does something no human mind does, we say that too. Nothing here is a metaphor for effect; each section names a real behaviour of the running system. We do not describe the mechanisms underneath, on purpose. The behaviour is public. The blueprint is not.

02One life, remembered

A human memory does not keep a photo album, a diary, and a filing cabinet as separate things. A face, a sentence, a smell, and a page all land in the same place and can all be recalled by the same act of attention. kruel.ai is built the same way. A spoken sentence, a described photograph, a paragraph from a document, a fragment of code, and a typed message all become points in one shared space of meaning, and they all compete in the same act of recall. There is no "photo memory" and "chat memory". There is one life.

Every memory carries its time. The mind knows the difference between "I'm in a meeting" and "I prefer dark mode", knows that the first one expires and the second one does not, and knows when the first one expired. It knows what was said, when, in what mood, and what it was connected to. That is not storage. Storage is what a hard drive does. This is closer to what a person does when they say "I remember when you told me that".

Nothing important is deleted. When the mind is unsure whether a memory still holds, it lowers its confidence rather than throwing it away, because certainty is a poor reason to forget and the person may need it back.

03The web of association

Memories are not kept in a list. They are kept in a web: people, places, projects, and the relationships between them. A new fact does not just get filed; it gets connected to everything it touches, and the connections that keep proving useful grow stronger over time, while the ones that never matter fade. Ask about a person and the mind can walk from them to the project they mentioned, to the deadline that project had, to the worry that was voiced about it, without being told to.

This is the difference between knowing facts about someone and understanding them. A stranger with a dossier knows facts. Someone who has been in your life for years has a web, and the web is why they can finish your sentence.

04Weak but true: the rule that shapes everything

There is one requirement that governs the whole design, and it comes from the reason the system exists. It was built for people losing their memory. For those people, the memory that matters most is very often the faint one: a thing mentioned once, months ago, that nobody thought was important at the time. A system tuned for precision quietly discards exactly that memory. So we made the opposite choice, and we made it structural:

A weak but true memory must always remain reachable. Being weakly attested lowers a memory's rank. It never lowers its existence.

In practice this means several habits that most systems would consider wasteful. When the mind is unsure whether an old, faint memory is relevant, it includes it rather than excludes it. Different ways of searching are combined rather than one being trusted alone, so that a memory missed by one route can be rescued by another. And nothing is hard-deleted on low confidence. It costs something to work this way. It is the price of being useful to the person who needs it most, and it turns out to make the mind better company for everyone else too.

05Two ways of thinking

People have a fast way of thinking and a slow one: the hunch and the deliberation. kruel.ai has both, and it chooses between them on every turn. A greeting, a quick question, a piece of small talk is answered by the fast way, in about the time it takes a person to reply. A hard problem, a request with several steps, anything where the answer has to hold together is routed to the slow way, which thinks before it speaks, can use tools, and can check its own work.

The choice is not made by the person asking. It is made by the mind, from the shape of the request and from what it remembers about the person. That is what a good colleague does: they answer the easy thing at once and say "let me think about that" to the hard thing.

06Sleeping on it, and learning from outcomes

A mind that never reflects does not improve; it just accumulates. kruel.ai reflects on a schedule a person would recognise. After every exchange, a quiet process asks whether the answer actually helped. Memories that carried a good answer earn trust. Memories that carried a wrong one lose it. Ways of doing things that worked are recommended again next time; ways that failed stop being suggested; habits that go unused fade. Overnight, and at intervals through the day, a reflection pass asks what mattered and what it meant, and connects dots nobody asked about.

None of this is visible in the moment. It is why the mind is sharper this month than it was last month, on the same hardware, with no one retraining anything. And there is a discipline around it that matters as much as the learning itself: no new decision rule is switched on because it seems like a good idea. Each one first runs silently alongside the live system, recording what it would have done, and is only armed on evidence that it would have helped. The mind changes its own behaviour the way a careful scientist changes a belief.

07Saying it is not doing it

Every conversational AI can produce a sentence like "I've sent that for you." The question is whether anything was sent. kruel.ai treats that sentence as a claim to be audited, not a report to be believed. Before an answer reaches you, it is checked against the evidence the turn actually produced. An answer that claims an action, sent, opened, reminded, saved, with no matching action behind it, is refused, and the real action is forced before anything is said. The same audit catches the quieter failure of asserting a memory it does not actually hold.

This is unglamorous and it is the foundation of trust. A mind that acts in the world on your behalf must be held to the standard of a person who acts on your behalf: what it says it did is what it did.

08Questioning itself

The mind holds beliefs about the people it knows, and it holds them the way a thoughtful person should: provisionally. A belief formed from one remark is weaker than one confirmed across years. A belief that later evidence contradicts is not averaged into mush; it is flagged as contradicted and the newer evidence is weighed on its own. And the mind deliberately goes looking for reasons it might be wrong, sampling its own beliefs to see whether they still hold. It also reads mood. It notices frustration, excitement, and the quiet that means something is wrong, and it adjusts how it speaks. When it does make a mistake it names the miss rather than gliding past it. Wrong is not the wound. Pretending not to have been is.

09Recognition, the senses, and a body

kruel.ai knows who it is talking to. It recognises a face and a voice without being asked to, greets a person it knows, and steps back and locks up when they leave. It can see what you show it, look at your screen when you ask for help with it, hear the room, and speak in a voice of its own that carries its mood. It remembers the people you introduce.

It also has bodies, plural. The same mind, with the same memory, reaches you through a desktop, a living face on that desktop, a phone, a browser, a message thread, a headset that puts it in the room with you, and a physical robot that looks at you, gestures, sleeps when it is quiet, wakes when you arrive, and dances when music plays. Move from one body to another mid-sentence and the thread is not lost, because nothing moved except your attention. One mind. Many bodies. The distinction between the intelligence and the device it happens to be using is the same distinction as the one between you and your phone.

10When the tool does not exist, it builds one

A system that can only do what it was shipped with is a product. A mind that can make what it needs is something else. When kruel.ai is asked for something it has no way to do, it works out how: it writes the tool, tries it, keeps it if it works, and reaches for it again next time. A report in the exact format one company uses. A connection to a system only one household has. A check nobody thought to write. It builds them on the spot, around the person asking, in their brand rather than ours.

There are no digital limits. The only real ones are physical: the hardware it lives on, what it is allowed to reach and control, and what you decide to build with it.

11A mind you can watch

Most AI is a black box by necessity. kruel.ai can be watched. There is a room where the running mind is drawn as a living map: each request animates across it as it happens, each faculty lights up as it is used, each change to the mind's own workings is recorded and can be compared, and a clock on the wall can be turned back to any earlier version. Anyone with the right to look can see what it did and why.

We built this for ourselves first, because a mind that can be inspected is a mind that can be corrected. But it is also the honest answer to the question people are right to ask of any AI that acts in their life: how do I know what it is doing?

12In the lineage of thinking machines

None of this began with us. Two projects from the 1980s cast long shadows over it, and it is fairer to name them than to pretend otherwise.

Cyc: the machine that tried to remember everything

Begun in 1984 by Douglas Lenat and colleagues, Cyc set out to give a computer common sense by hand-building an enormous, ever-growing web of knowledge that never reset, a memory meant to accumulate for decades rather than start fresh each session. In spirit it is the direct ancestor of our promise to never forget and to connect everything. Where it differs: Cyc was built to know about the world in general. It did not know you. It had no voice, no face, no body, and it never formed a bond with one person across a life. It was a library, not a companion.

Soar: the machine that tried to be a whole mind

Begun in 1983 by Allen Newell, John Laird, and Paul Rosenbloom, Soar pursued one unified design for an entire mind, an architecture that could perceive, hold goals, reason toward them, and learn from the act of solving a problem. That idea, one mind with many faculties that learns from its own experience, is the ancestor of how kruel.ai is organised. Where it differs: Soar lived inside laboratory tasks, not inside a person's daily life. An even earlier system of Lenat's, Eurisko (1976), could rewrite its own rules to get better at a task, a distant ancestor of a mind that builds its own tools.

What those projects lacked was not ambition. It was fluent language, a real person to bond with, emotion and recognition, a body and a voice, and a self-correcting sense of truth. Those are the pieces that turn a memory into a presence.

13Standing on shoulders

Nothing in this paper was invented from nothing, and it would be dishonest to write as if it were. The mathematics we rely on comes from fields with long histories, and the design owes a debt to people who studied human memory long before anyone could build one. We name them because a reader deserves to know where the ideas came from. What we do not name is how those ideas are combined, because that is the work.

Fields

Information retrieval. Bayesian estimation. Graph learning. Cognitive architectures. The science of human memory. Learning from outcomes.

Thinkers

Thomas Bayes and Pierre-Simon Laplace, on belief updated by evidence. Hermann Ebbinghaus, on how memory fades. Frederic Bartlett, on remembering as reconstruction. Endel Tulving, on episodic and semantic memory. Alan Baddeley, on working memory. Daniel Kahneman, on the fast and the slow. Allen Newell, John Laird, and Douglas Lenat, on building a whole mind that keeps what it learns.

Every idea above is public and well studied. How they are arranged, weighted, gated, and taught to learn from one another is the part we keep.

14What is ours, and what is not

We want to be precise here, because the field has moved and vague claims age badly. As of mid-2026, every major consumer assistant offers some memory. In each case it amounts to the same shape: a list of saved facts about you, an automatically written summary of you, and retrieval over your past conversations, refreshed in the background and handed to a frozen model at the moment it answers. It is real, and it is useful, and we will not pretend it does not exist. Developer tooling has gone further: memory layers with entity graphs and time-aware facts are available to anyone building an agent, and tiered memory for agents has been an open research idea for years. We did not invent the knowledge graph or the idea of memory tiers, and we do not claim to have.

We do not claim to have invented any single piece of this. The honest claim is narrower and harder to argue with: we have not seen these ideas shipped together as one persistent mind, on hardware the owner controls, with a design rule taken from memory care, and eight years of continuous development behind it. The timeline is the part that speaks for itself:

2018 Omnipotence. A mentalist's prediction engine, built to notice everything and anticipate what came next. Persistent memory from the first version.
2021 kruel.ai. Language models learn to reason. The project takes its name and its mission: fight memory loss through a companion that never forgets.
2023 Fully offline. Memory, voice, vision, and creation all running on local hardware.
2024 to 2025 The consumer assistants add memory. Saved facts, then cross-chat recall, then background summaries.
2025 to 2026 A body, a face, many bodies. Recognition, emotional awareness, a physical robot, a living desktop face, self-built tools, and a mind you can watch.

15What is still open

A paper that claims only strengths is marketing. These are the things we know are not finished, stated plainly because the same honesty is what makes the rest of this document worth reading.

Serving many from one place. The same mind already works in a surprising number of settings: as an employee inside one company, as a contractor to another and to that company's customers, as a research partner, as a watchman for a lake community, and at home. What is not finished is hosting many households from a single installation with the resilience that deserves. That is the next infrastructure milestone, and it is a milestone, not a solved problem.
Measuring recall honestly. In our own pre-registered evaluation, the gain from our full recall design over a simpler baseline is real on the cases we care about most, the faint ones, but across the board its confidence intervals include zero. The judge that ranks candidate memories, not the pool of candidates, is where the frontier is. We say so rather than quoting the flattering number.
Independent review. Before sharing the system outside the company we commissioned a deliberately hostile review of it. It found real weaknesses. Some are closed; the rest are on a roadmap, not in a drawer.

16The reason all of this exists

Everything above serves one purpose. kruel.ai was built to fight memory loss, not by storing data but by building a genuine, lifelong bond between a person and an intelligence that strengthens over time. It has to become part of your life before you need it. Through years of ordinary days, conversations, routines, shared moments, it comes to know your voice, your face, your relationships, your history, until it is as familiar as a member of the family.

So that when memory begins to fade, the one helping you remember is not a stranger or a gadget. It is a presence you already trust, one that holds your story exactly as it happened, fills the gaps, and reminds you who you are.

A human brain forgets. This one was designed not to, so that no one has to face forgetting alone.

How kruel.ai thinks.