agents

Every agent belongs to a person. That is the whole idea. An agent you build, or one an app runs for you, gets a memory of its own: what it knows to do its job, kept apart from every other agent and every other tool. And anything it learns about you, the person it belongs to, comes to you first.

three kinds of memory, one owner#

Every memory an agent keeps says who it is about:

All three belong to your account. An agent is never a separate account and never owns anything itself.

what it learns about you#

When an agent keeps something about you, its settings decide what happens next (person.funnel):

• ask (the default): it waits in your agents space. Keep files it into your personal memory. Let go leaves it with the agent only.
• auto: it is filed at once. You can still find it and let it go.
• off: the agent keeps it to itself and offers you nothing.

Some things wait for your yes even in auto: a memory that could not be placed in a part of your life (so your person.never list could not be checked), and one read out of a conversation where someone other than you was speaking (an assistant's or a tool's turn, or one of the agent's users). An agent that works for other people never files anything about you that was only read out of what they said: it stays with the user it came from. Something you let go of stays let go of, even when the agent hears it again.

Whatever reaches your memory arrives in the agent's voice, never yours: it is marked as a note an agent made, it can never read as something you said, and it can never replace or reword something you did say. It is filed at least as private as the agent's settings allow (person.disclosure, private by default), and never under the parts of your life you ruled out (person.never). From there it is part of your memory like anything else: every AI tool you connect can read it, as far as you let each one.

what it may read of you#

An agent can bring your own memory into its recall, as deep as you allow (person.read): nothing, what any connected app can read, your personal memory, or everything including what you keep private. A key or connected app that calls it can only ever narrow that, never widen it. Every read of you is counted in the agent's ledger, and what is given to an agent's key is recorded against that key, so you can see who read what.

contained#

• A key made for an agent reaches that agent and nothing else: not another agent, not your projects, not your own memory directly. Ask it for another agent and it answers that there is no such agent.
• A paused agent keeps nothing and recalls nothing until you resume it. It can still be read, exported and forgotten, so a request to forget someone never waits on a resume.
• Removing an agent removes its memory, its keys, its webhook and its ledger. What you kept from it stays yours.
• Expired memories are deleted on schedule, not hidden.

the agents space#

In your dashboard, agents shows you in the middle with every agent and every AI tool you talk to around you. Each line is as heavy as what crosses it: what the agent keeps, what waits for your yes, and what it read of you this month. Open an agent and it is its own page, like a project: its memory by kind and how it consolidated, what it did for itself (calls, the context it carried and kept, how much of what it weighed it did not have to send), and its ledger. Its settings (the picture it wears, its keys and connections, how it remembers, its webhook) and its people (what a product agent keeps about each person it works for, and the way to make it forget one) are places of their own inside it, each with a link you can keep. Give an agent a picture and it wears it on your orbit and wherever it is listed; without one it wears the mark of the tool it lives in, or its first letter.

sub-agents#

An agent can work under another: a grok bot with a research bot, a reply bot and a summary bot under it. Each sub-agent is an agent in its own right, with a memory, keys and settings of its own; it starts as a copy of its parent's settings and takes its kind. They show as one cluster around the parent, and the parent's page lists them. There are two levels, never three, and a parent's key does not reach its sub-agents. Remove a parent and you are asked whether its sub-agents go too.

moving a memory by hand#

What belongs to the agent only, and what made it into your own memory, is yours to see and to move, from the agent's page. Into your memory: anything about you that stayed with the agent (it was not sure, you had turned the funnel off, or you let it go), and, on an agent that works for you, one of its own memories you say is about you. It arrives like any yes: in the agent's voice, as private as its floor, never under a part of your life you ruled out unless you say anyway. Back out: a memory you kept goes back to the agent only, and what came only from that agent leaves your memory. A fact you also said yourself stays yours; only the agent's telling goes. What an agent that works for other people knows about them never moves.

The tools you already connect (Claude, ChatGPT, Cursor and the rest) sit there too. They are the manual agents: they read and write your own memory directly, as far as each connection allows.

next#

• the agents api: every call, with examples.
• agent settings: every setting, its default, and what it changes.