What Nomi remembers

Nomi keeps what it learns about your brand between conversations, somewhere you can read it. Open Brand in the rail and pick Agent memory in its sidebar. It is a dropdown: it opens onto the memory's folders, such as episodes, preferences, products and, when there are some, lessons or skills, and a folder opens onto its memories. Click a memory and it opens as the page. Everything in there belongs to the brand you are in, so a second brand starts from nothing.

With no memory open, Agent memory shows what Nomi knows about the brand overall, with what your ads have shown above it.

What goes in it

Three different things sit side by side.

• What Nomi believes, filed by subject: a preference you stated, a correction you made, a fact about your business that should outlive the conversation. Nomi files each one under the subject it is about.
• What Nomi has learned, in lessons: one page per subject, such as your offer, your voice or the angles that work for one product. Each page says what holds now and, under History, the moments it learned it from. brief is the short version Nomi reads at the start of every conversation.
• A record of each day's work, one file per day, in episodes. It holds what you asked, what Nomi ran, what landed and what it learned from each turn, never what a web page or a search returned. You see a turn there only when you can open its conversation, so a teammate's private conversation stays out of what you read.

What the ads have shown

The overview of Agent memory carries a second list Nomi did not write from anything you said: what it measured. Once an ad account is connected, it compares that brand's ads against each other every night and keeps what holds up, with the number of ads behind it.

It reads every ad the account ran, not only the ones Nomi made, so a brand arriving with a year of an agency's work has evidence from the first night.

Nothing there needs answering. If one is wrong, Stop using it and Nomi drops it from the next turn.

How something gets remembered

Nomi writes a belief mid-turn, without asking, and its working steps say Remembered something with the sentence it kept. That sentence has to be your own words, copied out of a message you typed: anything it wrote itself, or read off a page, is refused before it can be stored.

When it is unsure something is worth keeping, or when you told it to remember something in so many words, it asks instead. The question takes the place of the message field: Remember or Don't. Your answer stays in the conversation, with a link to what it kept.

There is no button for adding one by hand: tell Nomi to remember it, and it asks first.

How Nomi learns

After each message you send, Nomi notes in a line or two anything worth knowing next time: something you decided, turned down, corrected or prefer, or a fact about your business. Most messages hold nothing like that, and then it notes nothing. It works only from what you wrote and what it said back, never from a page it read or a competitor's ad.

Every few notes, it files them into its lessons: onto the page about that subject, or onto a new one when nothing fits yet, and it rewrites brief. None of this costs you credits.

Correcting it

Open a memory. Its bar shows the file name, Made by Nomi when Nomi wrote it, and when it was last updated. On the right, a two-way switch flips between the styled text and the raw text, the second of its two glyphs being the code one. The raw view is where you correct a memory: type in it and it saves when you click away. There is no Edit button.

The ⋯ beside the switch holds Pin to the top, which reads Unpin once it is pinned and gives the memory a mark in the sidebar's list, Move to followed by a folder's name, and Forget it.

A day's record, in episodes, is the exception: it is rewritten from what happened every time you work, so its bar says Rewritten as you work, there is nothing to edit in it, and its ⋯ holds only Forget it. If a day says something you disagree with, the belief behind it is what to correct.

A lesson you edit or pin keeps your wording: Nomi still adds to its History, but no longer rewrites what the page says holds now. A lesson you forget stays forgotten, and Nomi does not write that page again. Forget brief and Nomi writes a fresh one on its next pass.

Only a brand admin can change any of this. Everyone on the brand can read it, except that a day's record shows each person only the conversations they can open.

What Nomi does with it

Nomi starts every conversation from its brief and looks up the rest when it needs it, so a new conversation still starts knowing roughly where you left off and you should not have to tell it the same thing twice. Its working steps show when it is checking its memory.

It searches everything it keeps about the brand at once: its memories and lessons, the day's record, what was said in your other conversations that you can open, your products, your personas and your brand details. When a result is worth more than a few lines, it opens it whole. The search matches meaning as well as words, so a question in Slovenian finds a note written in English.

What you say in the conversation always wins over what it learned before.