Harlyn
LEARN · MEMORY, DONE CAREFULLY

Why an AI delegate needs a standing brief, not a memory

Every AI product now claims to learn about you. For a meeting delegate, a tool that speaks in your name to people who matter to you, that claim needs a much more careful answer than "yes." The useful question is not whether it learns. It is what it does with what it learned, and who decided that.

This piece is about the difference between a memory and a standing brief, and why a delegate should have the second.

The problem a delegate has that a note-taker does not

A note-taker's output is a document you read later. If it remembers something wrong, you see the error before anyone else does. A delegate's output is speech, in a live room, on your behalf. If it remembers something wrong, a customer hears it first.

That asymmetry changes what "learning" is allowed to mean. A delegate that silently absorbs everything it hears and folds it into future answers is not more helpful. It is a liability with a good memory. The approval you gave once, in one negotiation, for one customer, becomes a standing policy nobody wrote down. The number you mentioned as a hypothetical becomes a fact. The thing you said to end a conversation becomes the thing it repeats.

Memory versus a standing brief

Chat memory, the kind assistants build up from your conversations, is unscoped and unconfirmed. It accumulates. You rarely know what is in it, you did not approve each item, and it applies everywhere.

A standing brief is the opposite on all three counts. It is a short, readable list of positions the delegate may state, context it knows but never says, and rules for how it behaves. Every entry was confirmed by you. Every entry has a scope: one company, or every meeting. And every entry has an expiry, so a position you confirmed in spring does not survive unexamined into autumn.

The rule that makes this safe is simple enough to fit on one line: learned means suggested; confirmed means speakable. The delegate may notice patterns and propose. It may not act on a proposal until you accept it.

How it grows without you writing a manual

Nobody wants to fill out a questionnaire about their positions. A standing brief has to grow from work you are already doing, or it stays empty.

There are two natural moments. The first is when the delegate defers a question to you and you answer it, from Slack or from the dashboard, and your reply is spoken into the room. That answer is already in your words. Keeping it is one click: for this company only, or for every meeting. The second is when the room keeps asking the same thing. If the same question has been deferred to you in two or more meetings, the delegate proposes your latest answer as a standing position, with the evidence attached: how many times, and when it last came up. You edit the wording, pick the scope, and add it, or dismiss it and never see it again.

Both paths run through you. Neither involves the delegate deciding on its own that something is now true.

What a buyer should ask

Four questions separate a standing brief from a memory with good marketing.

  1. Can I read the whole thing? If the answer is a settings page you can scroll in under a minute, good. If the answer is "it learns from context," you cannot audit what it will say.
  2. Did I confirm each entry? Suggestions are fine. Silent adoption is not.
  3. Is it scoped? A position that is true for one customer is often wrong for another. The brief should know the difference.
  4. Does it expire, and is its use logged? Positions go stale. And when the delegate walks into a meeting carrying standing entries, the meeting's own record should say which ones.

How Harlyn does it

We build Harlyn, and this is the design we shipped in September 2026. The standing brief lives in Settings as three short lists: positions, internal notes, and rules. It is prepended to every meeting's briefing. Entries come from your own edits, from answers you chose to keep after a live deferral, or from suggestions drawn from questions you were asked repeatedly and answered. Suggestions are never spoken. Confirmed entries expire after 90 days unless you reconfirm them, the morning digest reminds you which ones are about to, and each meeting's activity record lists the standing entries the delegate walked in with.

That last detail is the one we would push any vendor on. A delegate that learns should be able to show its work, meeting by meeting. Otherwise "it learns about you" is just a nicer way of saying you no longer know what it will say.

KEEP READING
LEARN · THE ETHICS OF SPEAKING FOR SOMEONE
What should an AI say in a meeting on your behalf?
LEARN · THE APPROVAL GATE
AI meeting agents that ask before they act: the case for the approval gate
LEARN · THE CATEGORY IN 2026
An AI meeting assistant that attends for you: what it can actually do in 2026
See a delegate handle a real meeting.
Harlyn is in private beta, free while we build with early users. Every claim on this page is shipped behavior.
Request beta access Talk to Harlyn live
© 2026 Harlyn · in the room, even when you're not.
Trust Privacy Terms