Knowledge Hub and Agent knowledge

Fix your bio once and every agent that uses it is correct.

Your material belongs to the account, not to one agent. A source you add while working on your Support Agent can feed your public twin too, and neither of them holds a second copy of it. There are two screens, and the difference between them is the whole idea.

Knowledge Hub — the account’s library#

The page is your list of sources. Add a source, at the top right, opens the add form; Back to sources in the same place takes you back. An agent’s own knowledge page has the same pair. An empty hub shows the six ways in as doors, each opening the add form on that way.

Sources#

Everything the account holds, grouped by where it came from — the site, then the section of it — exactly as an agent’s own page groups its material, and exactly as the picker grouped your pages when you imported them. Each group says how many sources and how many words it holds, biggest first.

Above the list, kind cards — Web pages, Video, Files, Notes, and Written by hiy when there is any — say how many of each you hold, and each carries a + that opens Add on that kind’s way in. Press a card to narrow the list to that kind; press it again to clear. A kind at zero can’t narrow (there’s nothing to show) but still offers its +, so the cards also tell you what you have none of.

Search matches the title, the type, the address and the agent a source was added on. Search and the cards open every group that still has something in it, so nothing matching stays hidden behind a closed heading.

Each row says:

• what it is — the type badge, the first line of what hiy read from it, and the word count;
• where it came from — the agent it was added on, or Added here for something added straight to the Hub, which belongs to no agent at all. An agent’s name links to its own knowledge page, which is where you’d remove a source that was added there;
• which agents use it — this is the part no other screen can show you. A source can feed one agent, several, or none at all;
• whether it’s confidential — see below.

Removing a source added on an agent isn’t here: it lives on that agent, next to the warning about what removal means. Something added straight to the Hub has no agent to remove it from, so its Remove is here — and if an agent is using it, that agent has to stop first.

Add — to the hub, and to no agent yet#

The add form is the same six-way form as an agent’s own — paste text, an assistant’s memory, a file, a web page, a YouTube link, a LinkedIn export — with one difference, and it’s the point of the screen: what you add here goes into your library, not onto an agent.

Nothing uses it until you say so. Open the agent that should have it, press its Add a source, and pick it out of the hub — search, tick, add. From then on the hub shows that source tagged with that agent.

It works the other way round too, and always has: anything you add on an agent appears in the hub straight away, tagged with the agent it was added to. The hub is the whole picture; the agents are what draws from it. Adding a web address the hub already holds — on any agent, or through a bulk import — puts that agent on the existing source rather than reading the page again.

How much the hub holds. On Free and Founding, what you add here has an allowance of its own — five times your plan’s words and sources — and the add form prints it above itself so you can see where you are. Material you add on an agent counts against your plan’s one allowance, shared by all your agents, and using a source on a second agent costs nothing — it isn’t copied. On Early Access there is no separate Hub allowance: the add form prints the one pool your agents and the Hub share. See Limits.

Who can change what. The Hub’s own material belongs to the whole team: anyone you’ve invited can add to it and take things out of it. Two limits. Marking something confidential is an owner or admin only, because it’s the one switch the database itself enforces downstream. And a source that was added on an agent stays that agent owner’s to change — team-wide editing covers what was added to the Hub, not everything visible in it. A colleague who owns no agent of their own also has no Hub screen to reach yet; they see the inbox they were invited to.

Removing something. A source in the hub that no agent uses is removed in one press. If an agent is using it, the press asks which you meant: take it off that agent, or remove it from the library. Take it off every agent first — a source disappearing from under an agent that answers from it would change what that agent knows without saying so.

Syncing a site#

Web pages change after you import them. Sync, on any site’s group, does two things in one press:

1. Re-reads every page you already hold from that site and replaces the text where the page changed. Your title stays yours — a rename you made is never overwritten.
2. Asks the site what else it has now, and shows anything you don’t already hold in a checklist. Nothing new is added unless you tick it.

You get a report either way: how many pages changed, how many are new, and any that couldn’t be read — with the reason, page by page. A page that fails is left exactly as it was; a sync never deletes anything.

Sync is offered on sites, and only for material with a web page behind it. Pasted text, uploads and videos have no address to go back to, and the panel says how many of the group it’s leaving alone.

Re-read, on a single row, is the same first half narrowed to one page. Use it when one article has changed and the rest of the site has not: it fetches that page, replaces its text, and tells you whether anything moved — “No change since it was last read”, or the number of words gained or lost. It never asks the site what else it has, so it offers you nothing new to tick; that is what the site’s Sync is for. Like Sync it does not train, and like Sync it is offered only on rows with a web page behind them — a pasted note has nothing to go back to. On Early Access, re-reading one article here spends the same once-a-week-per-site clock as that site’s own Sync — see Early Access re-reads. Every source records when it was last read — including a re-read that found nothing changed, which is the commonest outcome and still a read — so Last read is a fact about that row rather than something you have to remember doing.

Agent knowledge — one agent’s view#

The same list, from one agent’s point of view, in two groups:

• From Knowledge Hub — sources added on another agent that this one draws on. Each says which agent it came from, and stop using takes it off this agent only. The source stays in Knowledge Hub and every other agent that uses it is untouched.
• Added here — sources whose home is this agent. These are the ones you can rename, hold back, open, and remove. Each row carries an In answers switch: untick it and the source is held back in place, tick it and it’s back in — at the next train either way, and never a removal. Rows from Knowledge Hub don’t get the switch, because held back is a fact about the source and would silence it for every agent drawing on it; those rows have stop using instead.

On an account with one agent there’s nothing to come from, so the list stays flat until you have a second one.

In the newer app the page has four tabs — Sources, Lists, What it knows and Map (a personal twin’s only; see Lists and Mind) — listed under Knowledge in the agent’s side list, and Guardrails moved to the agent card’s sheet on its Overview. The older console has three tabs under the title — Sources, What it knows and Guardrails — with Add a source at the top right, and one thing under them, above every panel: Words used, stated once for the screen rather than beside every tab. It counts your whole workspace against the plan, because that’s what the limit counts; with more than one agent, the ? beside it says how much of that this agent holds. While the workspace is Paused it shows the counts and the Paused sentence instead of a bar.

Finding one page among four hundred#

Once you’ve imported a site, Added here stops being a list you can read top to bottom. So it isn’t one. Sources are grouped by where they came from — the site, then the section of it — and each group says how many sources and how many words it holds.

Groups are ordered biggest first, and a bar under each one shows its share of the level above, so the thing you have too much of is the first row you read. Open a group to see the sections inside it, or the pages themselves. A long group shows the first twenty pages and offers the rest in one press.

Above the tree there are two ways to narrow it:

• Search matches the title, the type and the address — so billing finds the section as well as the pages that mention it — and opens every group that still has something in it, so nothing matching stays hidden behind a closed heading.
• Kind cards — Web pages, Video, Files, Notes, Written by hiy — each with its count and a + into Add. Press one to filter to that kind, again to clear. Kind is deliberately a filter and not the grouping: “your 399 web pages” is not a heading you can navigate by. (Inside an open source, the list beside it keeps the compact chips — same filter, narrower rail.)

Both narrow the same tree at once, and a line under it says how many of your sources are showing. A short list — eight sources or fewer — keeps the plain list and has no search box, because there’s nothing there to find; the cards still appear, since three sources still have kinds and a +.

Guardrails#

The Guardrails tab (a section of the agent card’s sheet in the newer app) is what this agent will and won’t take on: every off-limits topic in full, the subjects it covers and what it does when asked something outside them, whether it’s been interviewed, and — for a Support or Team Agent — where handing over to a person is set.

It states them; it doesn’t edit them. Each group opens onto the current answer and a way through to Settings → Rules, which is the one place these are changed, so two screens can never disagree about what you last said. See What it answers for what each setting does.

Taking something from the library#

Use something from Knowledge Hub opens a checklist of everything the account holds that this agent isn’t already using. Tick what you want and press the button; nothing is copied, and it costs nothing against your source limit — the same source now feeds both agents.

What a shared source does when you change it#

Each agent builds its own index, so each one picks a change up at its own next train — Train now on that agent, or the overnight one if it has Train overnight switched on.

• Removing it removes it everywhere. It’s one source, not a copy per agent, so taking it out of the agent it was added on takes it out of all of them, and out of the hub as well — which is why remove asks first, in the row itself, naming the source, and says so on the row if it doesn’t go through. The hub and the other agents stop listing it at once; the row you pressed it on stays, marked Removing, until that agent’s next train takes it out of the answers.
• Holding it back works the same way: a held source is out of every agent’s index, not just the one you were looking at when you held it.
• Stop using, on an agent’s own page, is the one that isn’t global — it takes the source off that agent and leaves everyone else alone.

History#

History, beside Add a source at the right of an agent’s Knowledge page title, is that agent’s own recent edits — newest first, one line each: a correction, a source removed, a source detached, or a change to guardrails. Pick a row to read its detail.

Most kinds undo in one press, confirmed inline:

Undo replays the same action you’d take by hand, so it costs nothing extra and follows the same training rules as the edit it reverses.

In the new app#

Knowledge in the sidebar is the workspace library: every source in the workspace, whichever agent uses it. Pick an agent in the column beside the list, or with Agent on a phone, to see only the sources it uses; All agents shows everything.

The columns are Name, Type, Words, Used by, Last read and Confidential. Used by shows the first agent and how many more use the source; hover it for all of them. A source no agent uses says Not used by any agent. List | Map switches between the rows and each agent’s map.

An agent’s own Knowledge page shows only what it uses, under From library and Added here. Add from library picks sources the workspace already holds, Add new adds a fresh one, and Open in Knowledge on a row jumps to it in the library. A confidential source is never offered to an agent that answers the public.

What you agree to#

Importing the facts an assistant already stores about you — from ChatGPT, Claude or Gemini — asks you to tick one box first, and this is what it means. It is the only way in that asks: everything else you add is material you wrote or chose, and this is material something else wrote about you.

You are confirming that you have read the facts about to be stored, that they are about you and yours to share, and that hiy may keep them. They may include sensitive details — health, sex life or sexual orientation, politics, religion, finances or relationships — which is exactly why the box is there rather than a quiet default.

The second half matters more than the first. If your twin is public, a stranger who chats with it can be shown one of these facts as part of an answer, quoted, with that import’s title beside it as the source. Nothing is hidden by being a “memory”: a stored fact is material like any other, and material is what your twin answers from. Review the list before you tick, untick anything that shouldn’t be there, and remember that a source can be held back or removed afterwards.

Confidential sources#

Marking a source confidential keeps it to your Team Agent — the one kind only colleagues you invite can reach.

It’s blocked from your public twin, and from a Support Agent too. A Support Agent isn’t on a public page, but the people it answers are your customers’ users, not your colleagues, so confidential material has no business there.

What that means in practice:

• The picker on a public twin or a Support Agent doesn’t offer confidential sources at all, and says how many it’s holding back.
• You can’t mark a source confidential while one of those agents uses it. Take it off that agent first — the row tells you which one.
• Nothing confidential reaches such an agent’s answers, its summary, the way it learns to sound, or the suggested questions on its page. That check is on the read as well as the write, so it holds even if something writes around it.

Confidential is for material a Team Agent should answer from and the public never sees: internal notes, salary bands, unreleased plans. Your public twin keeps working; it just can’t see those rows.

Limits#

A source added on an agent is counted once, and your plan’s word and source limits count every agent’s material together — one allowance for the account, not one per agent. Something added straight to the Hub has no agent, so on Free and Founding it counts against the Hub’s own allowance instead — five times the plan’s words and sources. On Early Access there is no separate Hub allowance: Hub material shares the one pool with your agents’. Subscribing another agent to a source doesn’t spend anything — that would defeat the point of a library. The words meter on each agent’s Knowledge page counts the whole account, exactly as the limit does, and says beside it what that agent holds.