Skip to content
Clone-Me

How it works

Your clone learns your criteria in four steps. Nothing you haven't confirmed is ever used for a judgment.

  1. It takes notes as you work

    Your AI agent notes the moments you choose, correct the AI or give a reason, and the server reads across conversations to write principles for you to review.

    • Your AI agent doesn't write principles. It notes the moments you choose something, correct the AI, give a reason or stress a point: what you said, the situation, what you chose and turned down, and the reason you gave. It also keeps the context: what you were working on and what was at stake, what the AI had just proposed, the options on the table, other things you said about it, and what happened next.
    • Cases stay in the server's hidden notes for 10 days from when they were last seen, or 30 days once the server has sorted them and is holding them for the next sorting, up to 300 per person, and are then deleted. Cases that back a proposal or memory stay with it.
    • At most once an hour the server sorts the hidden notes, reading across conversations to write principles. A principle reaches your inbox only if it showed up in at least two separate conversations or carries at least two kinds of importance signal (corrected the AI, something was at stake, gave a reason, stressed it). There's no daily cap.
    • Sorting uses external AI. Like judgments and dreaming, it stays within the scope you agreed to at sign-up, and cases marked sensitive are never sent.
    • In Claude Code, your words in a case are checked against the conversation log on your computer to confirm you actually typed them. Which conversation a case came from is reported only as a hash of the conversation ID; the conversation itself is never sent.
    • Each proposal records its source (which connection, when) and whether it's something you said or an inference. Unknown dates or reasons are never made up.

  2. You confirm

    A proposal is not a memory until you confirm it.

    • In the inbox, you see each proposal's content, its source, why it came up and the cases behind it (your own words), then approve, edit or reject it. You can also select several and handle them at once.
    • The default is “not selected”. No answer is not approval.
    • Facts you stated and preferences the AI inferred follow the same rule.
    • When you reject, you can note whether it was a one-off or simply wrong. A similar principle comes back only after more new conversations pile up.

  3. Your AI puts it to use

    Your AI agent receives your context and criteria over MCP and uses them in its judgments.

    • The lookup tool (recall) returns confirmed information, goals, relationships, experience, likes and dislikes, with their conditions and sources.
    • The judgment tool (ask_me) returns which memories it relied on, and explains why it holds off when the evidence is weak.
    • It doesn't look at principles alone: it also uses current information such as deadlines, budget and staffing, and states the uncertainty when information is stale or conflicting. (Planned)
    • When nothing it knows would change the judgment, it asks you instead of guessing. Your AI agent notes your answer as a case, and it becomes a proposal once the server sorts it.
    • Your clone's answer is advice for your AI agent. It is not approval of any real-world action.

  4. Dreaming refines it

    Planned

    It refines only what it's sure of and asks whenever anything is unclear.

    • It reinforces a criterion automatically only when new evidence appears under the same conditions as one you already confirmed. The same event delivered several times counts once.
    • It reads the conditions behind each choice, not just the choice. Picking the cheaper option because the budget was tight doesn't turn into “always picks the cheapest option”.
    • The only thing dreaming does on its own is link supporting cases and counterexamples to confirmed criteria. Exceptions, conflicts, changed criteria, duplicates, outdated info, scope and new principles become owner requests (a question with options) in your inbox, and the answer you choose is applied right away.
    • Changes are kept so you can compare before and after, and edits you made in the meantime are never overwritten.

When there's no basis, it asks and waits

Your clone doesn't pretend to know what it doesn't. When no memory applies or memories conflict, it holds off and says why. Your AI agent then asks you directly.

When there's no basis

Tools used: ask_meremember

A new kind of choice, so your clone holds off and your AI agent asks you.

  1. You: Add a license to the library we're publishing.

  2. CloneChecked with your clone: ask_meOn hold

    There are no confirmed memories about choosing a license. Question for the owner: MIT or Apache-2.0, and why?

  3. AI agent

    I don't have a criterion for this, so I'll ask. MIT or Apache-2.0?

  4. You: Apache-2.0, because of the patent clause.

  5. Then: Your AI agent notes this choice and its reason as a case (remember). When the server sees the same criterion recur or judges it important, it raises a proposed principle; once you approve it in the inbox, it becomes evidence for future judgments.

Stop re-explaining. Work with an AI that knows your criteria.

Sign-up is invite-only for now. Leave your email and we'll send an invite as soon as a spot opens.

Join the waitlistAlready invited? Log in