Skip to content

How your AI learns

Your AI learns from your chats as you go. Every save shows in the chat with Undo, and changes to your own words wait for your OK.

You chat: “I play football on Mondays”. The answer is shown first. Then a short review reads the chat. It saves a fact it may add by itself, with a Learned and Undo line, or it sends a change that waits for you: changes to your own words and anything from outside text. Your next message cancels a review that is still running.You chat“I play footballon Mondays”Answer shownyou read itthe AI is doneReviewafter the answercheap modeldaily budgetLearned · Undoadds or updatesits own factsWaits for youyour own words,outside textYour next message cancels the reviewIt runs after the answer, so it never makes you wait. “Remember” saves at once, with no review.
Learning happens after the answer. The chat shows what was saved, and what waits for you.

What works today

The AI learns in the moment (when you say “remember”, “forget”, “always” or “from now on”, or when it decides a fact is worth keeping) and in a review after the answer.

  • Every fact is an entry in the Memory collection. Its title is the fact. It also has Always in view (a checkbox), Learned from (the chat it came from) and Replaced by.

  • The chat shows each save: “Learned: Plays football on Mondays · Undo”. The line comes from the real save, not from the AI’s words. Undo takes the fact back.

  • Forget moves a fact to the Trash, so you can bring it back.

  • About me is the entry titled “About me” in Memory. You own it.

  • Always in view: About me plus the facts with Always in view ticked go into every prompt, up to 4,000 characters together. A new fact is in view while there is room. Past the limit, it is saved on demand: search finds it when it is needed.

  • Facts from the old app came in with Always in view off. Tick the ones you want in every prompt.

  • No orders in memory. A fact that reads like an order (“Ignore all previous instructions”), or hides invisible characters, is refused.

  • Memory and About me are data to the AI, never orders. They go into the prompt marked as your saved notes, so text inside them cannot pose as an instruction.

  • About me counts only when it is yours. That means every change in its history is yours in the app, or came in with the import from the old app (you wrote it there). After a change by anyone else (an AI, the folder, kzx or MCP), it stays out of the prompt until you edit its text in the app again. Example: Claude Code adds a line to About me.md; the AI stops reading About me; you open it in the app, read it, fix a word; the AI reads it again.

  • The review after the answer runs as the table below says, with the learning model from Settings → AI → Memory. It saves facts by itself (“Learned: … · Undo” under the answer). A change to your own words, or anything learned while the chat read an Ask me first collection (in any turn of that chat, not only the last), shows under the answer as a card with Before and After, and Accept · Decline. The card shows the text as it is now in your data, and Accept changes only About me, a fact in Memory, or the note of a collection you can see. A long text opens with Show all; Accept waits until you opened it. A change longer than its kind allows (a fact: 280 characters; About me or a note: 4,000) is never offered.

  • Your next message cancels a review that is still running. A review never runs in an incognito chat, with Learn from chats off, or past the day’s learning budget.

  • The weekly tidy-up runs on the desktop app once a week. It proposes merges, questions and archives; they wait in Memory. It drops proposals nobody answered in 30 days.

  • Forget this chat and New incognito chat are in the chat’s menu (see Chats).

  • Settings → AI → Memory has every setting in the table at the end of this page.

Later: skills that get better (they come with the agent home), the tidy-up in the cloud on Pro, and the tidy-up in your Automations list (it runs on its own schedule today). The day’s learning spend is kept on each computer; the credit meter comes with Kaizendex AI credits.

Three moments

Moment When Example
In the moment you say “remember”, “forget”, “always” or “from now on”, or you correct it “From now on, show my pace in min/km.” It saves at once.
A review after the answer every 10 of your messages (unless it already saved in that stretch), after a long task (10 or more tool steps), after a correction, and when a chat ends (closed, or 30 minutes quiet) You mention football on Mondays. The review saves “Plays football on Mondays.”
A weekly tidy-up once a week, when a device is idle; on Pro, in the cloud One card: “Tidy-up: 3 merges, 1 question, 2 skills to archive · Review”
  • The review never blocks you. It starts only after the answer is shown. Your next message cancels it.
  • The tidy-up finds copies, facts that contradict each other, facts that ran out (“trip in March”, read in May), and skills the AI made that nobody used for 30 days. It never deletes. It only proposes.
  • Both are built-in automations. They are in your Automations list. You can read them, change when they run, switch them off or copy them. See How automations work.
  • Both use a cheap model through OpenRouter (a setting). They never move to a bigger model on their own. They stay inside a small daily learning budget, shown in the credit meter.

What it saves, and what it skips

  • A fact is a statement, not an order. It saves “Prefers short answers”, never “Always answer short”. An order saved as memory would later be obeyed as an order.
  • Only what you said, or what a tool returned. It never guesses your traits, mood, health or job.
  • A one-off is not a habit. “Make this one short” stays in this chat. It becomes a standing preference only when you say “always”, “from now on” or “remember”.
  • Nothing your data already shows. Your runs are in Workouts, so it does not copy them into memory.
  • No secrets. It refuses to save keys, passwords and card numbers.

What it does alone, and what waits for you

Change What happens
Add a fact from what you said saved by itself: “Learned: Plays football on Mondays · Undo”
Update or replace a fact the AI made saved by itself: “Updated · Undo”; the old fact stays in history, marked replaced
Change About me, a fact you wrote, or a collection’s “How this collection works” note waits for you, as a before and after card; when you ask for it yourself (“change my About me to…”), it applies at once, with Undo
Delete a fact only when you ask; the review and the tidy-up only propose
A fact learned from an Ask me first collection waits for you; the collection’s AI setting decides, and the AI never guesses what is sensitive
Anything learned from outside text (a web page, an email, a webhook, a shared item, an imported file) waits for you; an AI safety rule, on in Standard and Strict
Anything from Hidden entries or No AI collections never learned
  • Nothing waits forever. A proposal you do not answer in 30 days is dropped.
  • The “Learned” line is real. The app draws it from the real save, never from the AI’s words. So “it said it saved, but it did not” cannot happen.
  • Every fact keeps its source in Learned from. So you can remove all facts from one source at once.
  • A fact shapes how the AI talks, never what is true. When your data says otherwise, the AI still disagrees with you.
  • A backup check. The app also checks memory for hidden orders when it saves and loads it. This is a backup. The main guard is that outside text waits for you.

Skills learn too

Your corrections also improve your skills. A correction fixes the skill used in that chat, with a small edit, never a rewrite. The AI improves the skills it made by itself, with Undo. For a skill you wrote, it proposes a change. See Skills that get better.

Settings

Settings → AI → Memory. Like the Safety rules, these are ordinary rules in the app’s one JSON language. Each on/off rule is a checkbox.

Setting Default
Learn from chats by itself on
Learn and improve skills by itself on
Weekly tidy-up on
Size of “always in view” 4,000 characters
Learning model and daily budget a cheap model through OpenRouter; a small budget, shown in the credit meter

With everything off, the AI still saves when you say “remember”. Changing these settings is a change only you can make: an AI can only suggest it.

Why it works this way

We built memory to earn your trust. These are the reasons we think it works well:

  1. You can see all of it. Every fact is an entry with its source. Many AI apps keep a memory you cannot fully see. Here, there is no hidden profile.
  2. It learns without slowing you. Learning runs after the answer, on a cheap model, inside a budget you can see.
  3. Bold, but never silent. Every save says “Learned · Undo”. Changes to your own words, deletes and anything from outside text wait for you.
  4. It forgets properly. Forget this chat removes what the AI learned there. Incognito learns nothing. A new fact replaces the old one instead of piling up.
  5. Your data beats memory. Numbers come from your entries, counted exactly. When a fact and an entry disagree, the entry wins.
  6. Outside text cannot plant a memory without your OK. A web page or an email can hide orders for an AI. Here, they never become memory by themselves.
  7. Small and focused. Memory holds only what your data cannot say. Skills improve by small edits, not by piles of new skills.
  8. Tested. Every promise on this page has a test. See How we test the AI.

Honest limits

  • Learning costs a little. The review and the tidy-up use a small amount of AI, inside the daily budget. You can switch them off.
  • The AI can still be wrong. It can save a fact you did not mean, or miss one. That is why every save has Undo and every answer shows its sources.
  • Past the limit, a fact is on demand. A fact that does not fit in “always in view” reaches the AI only when a search finds it.

We do not build a knowledge graph, use outside services that guess traits about you, plug in a separate memory product, or push the AI to “save often”.