--- title: "What it means for an assistant to know you best" url: "https://kyberbot.com/blog/an-assistant-that-knows-you" description: "Knowing you is not storing everything you said. It is a memory you can read, correct and delete, kept where you hold the keys. Here is how to test yours." --- # What it means for an assistant to know you best October 8, 2026·4 min read·Editorial team A pebble-shaped assistant beside an open box of memory cartridges, one taken out, with a hardware key next to it On Monday you tell your assistant that your sister has surgery on Thursday and you will be at the hospital all afternoon. On Thursday morning it books a two o’clock call with a client, because the slot in your calendar was free. Nothing was lost. The sentence is still there, somewhere in Monday’s conversation. The assistant had the information. It just did not know you. That gap, between having data about someone and knowing them, is the whole problem. Most assistants today are very good at the first and structurally bad at the second, and the reason is not that the models are weak. ## The real cause: memory was never designed, only accumulated An assistant that knows you needs a model of you: who matters, what is going on this week, what you always want and what you asked for once. Most systems do not build that. They keep a transcript, search it when asked, and hope the right line comes back at the right moment. A transcript is not a memory. It has no idea which facts are durable and which are temporary, which were a joke and which were a decision. There is a second cause, and it is not technical. The more an assistant knows about you, the more valuable that knowledge is to whoever holds it. When the memory sits on someone else’s servers, every improvement in how well the assistant knows you is also an improvement in how well a company knows you. So people, reasonably, tell it less, and it knows them less. The incentive and the design point in opposite directions. ## Where knowing you goes wrong **Memory as a transcript.** Everything you said, nothing you meant. The assistant can quote you, but it cannot tell that “Thursday afternoon” overrides “the calendar is open”. **Memory that never forgets.** You mention once, in a bad week, that you hate mornings. Six months later it is still declining nine o’clock meetings for you. Facts without a date and without an expiry turn into rules nobody chose. **Memory you cannot see.** If you cannot list what it believes about you, you cannot correct it. Wrong beliefs stay invisible until they cause something. **One memory per product.** Your calendar’s assistant does not know what your mail’s assistant knows, and neither knows what you told the speaker in the kitchen. You are known in fragments, by several companies, none of them completely. **Memory held by someone else.** Even a well-designed memory is a liability if it lives where you have no control: a retention policy you did not write, terms that can change, a breach you will read about later. We cover the business side of this in [Your data is not the product](https://kyberbot.com/blog/your-data-is-not-the-product). ## What knowing you should look like KyberBot is pre-launch; what follows is the design we are building it to. **Facts, not transcripts.** A memory item should read like a note a good assistant would keep: “Sister’s surgery, Thursday afternoon, not reachable. Told on Monday. Expires Friday.” Short, structured, sourced and dated. The conversation is where the fact came from, not what is stored. **Told and guessed, kept apart.** What you said and what the assistant inferred are different kinds of knowledge. Inferences should be marked as inferences, shown to you, and promoted to facts only when you confirm them. **Forgetting is a command.** Delete a fact, a person or a whole period, and it is gone from every device. If forgetting takes a support ticket, the memory is not yours. **Scoped when shared.** Most work does not need all of you. When KyberBot hands a task to an agent, the agent gets a brief, not the memory: the three facts the task needs and nothing else. How a swarm works from a brief rather than a biography is described in [Anatomy of a deep agent](https://kyberagent.com/blog/anatomy-of-a-deep-agent). **Kept where you hold the keys.** The long memory lives on hardware you own, typically a node at home, encrypted with keys you carry. [The home node](https://kyberdeck.com/blog/the-home-node) explains why the memory belongs on a shelf rather than in a data centre. Once what the assistant knows cannot be used against you, telling it more stops being a risk, and the design and the incentive finally point the same way. **One memory, many bodies.** The card on your lanyard, the band on your wrist and the robot in the hallway should all know the same you. That is a subject of its own: [One assistant, many bodies](https://kyberbot.com/blog/one-assistant-many-bodies). None of this needs a bigger model. It needs memory treated as a product in its own right, owned by the person it describes. It is also the reason Super Intelligence, when it arrives, [has to belong to the people who use it](https://kyberfusion.com/blog/why-si-has-to-belong-to-people): the more capable the system, the more it matters whose side its memory is on. ## A test you can run today Ask the assistant you use now four questions, and write down the answers. 1. **What do you know about me?** Not “I don’t store personal data”, but a list. If it cannot produce one, it either knows nothing or will not show you. 2. **Where is it kept, and who else can read it?** A specific answer, not a link to a policy. 3. **Forget one thing.** Pick a fact, ask it to forget, then ask about that fact tomorrow from another device. 4. **What did you guess?** Ask it to separate what you told it from what it inferred. If any answer is “I can’t show you that”, you have an assistant that holds data about you, not one that knows you. That difference is what we are building [KyberBot](https://kyberbot.com/) around: a memory you can read, correct and delete, on hardware you own, working for one person. - memory - personal assistant - privacy - trust ## Keep reading - [How an assistant hands work to a swarm](https://kyberbot.com/blog/how-an-assistant-hands-work-to-a-swarm) - [KyberBot in a robot: an assistant with mass](https://kyberbot.com/blog/kyberbot-in-a-robot)