Skip to main content
An agent can be in your private DM, in a room with you and a teammate, and in a room with your whole team. What it may recall when it answers depends on who can read the conversation it answers in: it only recalls what every one of those readers could already read.
Before opening Connie Settings: if pending Chief of Staff setup sends you back to chat, finish the setup cards there, then reopen Settings. Direct Settings links follow the same check. New workspaces use free credits, with no card or trial. Billing (?tab=plans) remains available during setup. See Settings access during setup.

What an agent uses where

Sharing an agent grants permission to use it, not access to its owner’s chats or saved memory. Each teammate opens a separate private DM; the owner cannot read that DM just because they own the agent. The agent’s saved memory remains owner-only, and a teammate’s conversation does not teach that shared memory. The agent’s purpose and standing rules still guide its work; keep private facts in memory or private conversation, not in instructions intended for everyone using it. See Share an agent. For example, Yann has two private rooms with Chief, A and B, and a room C with Chief and Alfie. In A, Chief may use what it learned in A, B and C: Yann can read all three. In C, Chief uses nothing from A or B, because Alfie can’t read them. If Hana joins C later, Chief still uses nothing from A or B in C. A room open to the whole workspace is read by its members, while a public link and a Slack channel also reach people outside your workspace. So in a room shared through a public link, an agent uses only what it learned in conversations that also have a public link, never what it learned in a room open only to your workspace, and in a Slack channel nothing it learned in Connie. A workspace guest reads a room open to the workspace only if they’re in it. So in a room a guest is in, an agent uses what it learned in a room open to the workspace only if that guest is in that room too, and the same goes for opening that room’s files or its work there. What an agent learns about how its owner likes to work is kept for conversations that only its owner reads, such as the owner’s DM with it. It learns there in two ways: it saves what its owner decides, approves or corrects in the same reply (see The playbook), and it reads its DM with its owner again in the background after each reply to keep its memory up to date. It doesn’t learn in the background from a DM with anyone else or from a room with other people; there, the agent works from the room’s own history and the decisions made in that room. An agent can read its complete memory, correct it, forget selected information, or clear it. Its memory tool is always available: nobody can turn it off for an agent or for a workspace. Ask it to show what it remembers, replace an outdated preference, or forget something. Its memory is one versioned summary document. It holds what the agent learned about how its owner likes to work; the decisions its owner made are kept one by one in its playbook (never also in its memory), and your personal memory holds facts about you. The agent says which one it used: “my playbook” or “my memory” for its own, “your Memory” for yours. Changes to the memory document require the current revision. The first save into an empty memory document replaces nothing and needs no confirmation; replacing or clearing it asks for confirmation on a Changes agent memory card that says which revision it saves, that the earlier revision is kept, and shows the Text to save in full. After a successful change, it can read back the saved result. A conflicting save changes nothing and requires reading the latest memory again. Forgotten information stays excluded from later automatic learning. An agent changes only its own memory, even when you own several agents. It cannot seed, replace or clear another agent’s memory. Ask that agent directly in your DM with it, or have the decision saved for all your agents in the workspace playbook, below. The same owner and audience rules apply: this does not make private memory readable in a shared room or in someone else’s DM.

The playbook: decisions your agents keep

When you approve, decide or correct something that should hold next time, such as approved copy, your ICP, pricing or a rule, your agent saves it in the same reply and tells you in one line. It doesn’t ask you to confirm: it is your own instruction. Each decision is kept in your words, with the day you made it and the conversation it came from, and your agent works from it every time after, without you repeating it.
  • For this conversation, for one agent, or for all of them. A decision you make in a room stays in that room unless your words make it general: “keep your updates to three bullets” applies in that room only. Say “always”, “from now on” or “in every chat” and it is kept for that agent in all its conversations. In your DM with an agent, where you manage it, a decision is kept for that agent in all its conversations, unless you limit it to that chat (“here”, “in this chat”, “only in this DM”). A decision your other agents need too is kept for the workspace, so copy you approved with one agent is known to your other agents: say “save this for all my agents” or “everywhere”. The agent’s reply says in one line where the decision applies and that it stands until you change it. A teammate’s agents keep to their own owner’s decisions.
  • Holds are not saved. An instruction about the work in front of you that ends when you say so, such as “don’t delete anything until I say so”, “for now, only use the Paris List” or “not yet”, is followed in that conversation and not saved as a decision. A standing rule such as “always wait until I approve the copy before sending” is still saved. When you later say “OK, delete it now” in the same conversation, your newer words apply; the hold never reaches your other conversations.
  • Changing your mind. Tell the agent the new decision. It follows what you say now and saves it in place of the old one, which is kept as history. You can also ask it to list what it has saved, or to retire a decision.
  • Saved once. A decision your agent saves for all your agents is kept only in the playbook, not also in your personal Memory: if the agent saved the same thing to your Memory earlier in that reply, that copy is taken out of your Memory (archived, so you can bring it back), and it adds nothing to your Memory that the decision already says. Its reply tells you where it kept it. A decision kept for one agent only, or for one Workflow, List or Object, leaves what the agent saved to your Memory in place, because your other agents still read your Memory. A decision your agent saves from your message is never added to your Memory in the background either, so retiring it takes it out of use and leaves no second copy behind. Anything else you said in that message, such as a preference of your own, is still saved to your Memory as usual, and so is anything you ask it to keep in your Memory when it saves no decision.
  • Not copied into the agent’s memory. An agent’s memory never repeats what a decision says: at most it notes that your decisions are kept in its playbook. When you retire a decision, or replace it with a new version, the memory of each of your agents that still repeated it is updated at once, as a new memory revision (the earlier revisions are kept), so the old words stop applying everywhere. An edit to an agent’s memory that repeats a decision is saved without that part.
  • Only your decisions. An agent saves decisions made by its owner, in a message they wrote. An authenticated request you post through Connie’s MCP or API can qualify when it is linked to your actual message in that Connie conversation. A request from someone else in a room, a message another agent sent on your behalf, a generated Task brief or comment wrapper, something said on a call, a routine or a background check saves nothing. This does not enable learning from Slack channels or external calls without your verified Connie message. The agent can still follow a participant’s request in that room; it just isn’t saved.
  • Private stays private. A decision is used in the conversation it was made in, in conversations only you read, and in a room where everyone could already read where you made it. A decision you made in a private DM is never shown to a teammate, in a shared room or through one of your agents answering them.
  • Size. Every reply starts from your newest decisions, up to about 8,000 characters of them; the agent looks up older ones when it needs them. One decision is at most 2,000 characters.
If you delete the message a decision came from, or Edit and resend past it, the decision is no longer used.

Learning from your Tasks

Your agents also learn how you work from the Tasks they work on, in the background, about every half hour: comments and changes people make on a Task the agent is assigned to or has worked on, such as “never include agencies in prospect lists”. Only what a member of your workspace wrote; never a guest’s comment, an agent’s own note, a deleted comment or a Task copied from Records. A change on its own, such as moving a Task to Done or approving it, teaches nothing: an agent learns only from words someone wrote, in a comment, a title or a description. What it learns from a Task is used only where everyone reading could read that Task: in that Task’s room, or in conversations only you read. A teammate’s Task comment is never kept as your decision. Learned items are added to your playbook, and your agent reads where each came from (the Task’s name). They appear in the agent’s Decisions section like your other decisions, with From Task and its name. Ask your agent to list or retire them, or select Retire, as for any decision. Each learning pass is charged as AI usage, like the rest of your agent’s learning. Your agents do not learn from your Inbox replies. Tell your agent what should hold next time, in its DM or a room, and it keeps it as a decision.

Versions, and what was made under each

Every decision has a version. When you change your mind, the agent saves the new decision as the next version of the old one, so v2 replaces v1 and v1 is kept as history. The new version stays where the old one was (the same agent, all your agents, or the same Workflow, List or Object) unless you say otherwise. When it saves the new version, the agent is told what was produced under the old one, such as replies, Pages it wrote, Workflows it edited and Workflow steps that ran, so it can offer to bring them up to date. Ask “what was made under the old opener?” to get the list.

Decisions for one Workflow, List or Object

Say “use this for the send Workflow” and the agent keeps the decision for that Workflow. Its AI and Agent steps then read the decision on every run, with its version, next to their own instructions. The same works for an Object: AI fields that run on its Records read the decision. A decision kept for a List is read by your agents whenever they work, not by AI fields. Only someone who can open the Workflow, List or Object can attach a decision to it. A Workflow step also reads decisions you made where everyone in the workspace can read them: in a room open to the workspace, or taken from a workspace Task. A decision you made in a private DM reaches a Workflow only if you attach it to that Workflow, because what a Workflow produces is read by everyone who can open it. A Workflow or Object shared outside the workspace, or one that a guest or someone who is not a member of the workspace can open, reads only the decisions attached to it. When an agent writes something inside a conversation, for example a message draft or a piece of text it asks a model to produce while it works in a room, the same rule as for its replies applies: it uses only the decisions that everyone who reads that conversation could already read. A decision attached to a Workflow, List or Object doesn’t change that. In a room with a guest or a teammate, your private decisions are left out; in your own DM with your agent, all of your decisions apply.

Decisions that already live in a Task, a Page or your Memory

If the decision already exists somewhere, such as a Task called “Locked opener” or your ICP in your Memory, tell the agent “that Task is the decision” or let it save the decision when it reads the Task. The decision then names where it came from. If the Task or Page is edited later, the agent sees that its source changed after the decision was saved, and can ask you or update the decision.

Decisions for one connected account: its voice and topics

When several people post or send from their own LinkedIn or email accounts, each account can keep its own decisions: how it writes and what it talks about. Say “for Maya’s LinkedIn: all lower case, short lines, end with a question; topics are hiring and onboarding” and the agent keeps that for Maya’s account. It works the same for any connected account.
  • Used whenever something is written as that account. A post or scheduled post for that account, a message, InMail, comment or reply sent from it, and an AI or Agent step that says it writes as that account all use the account’s decisions, and record which versions they used. If an agent drafted a post before it had seen them, it checks the draft against them before it posts or schedules it (nothing is posted until then). Checking a draft against the account’s recent posts also returns them.
  • Kept with the account, not with one person. Anyone who may act from the account (whoever connected it, and the people it is shared with) can save, change or retire its decisions by telling an agent, and they see the same ones. Someone who may not use the account never gets them, and neither does a room or shared chat where someone who reads it may not use the account. Reconnecting the account keeps them.
  • Other accounts are unaffected. An account with no decisions is written for exactly as before, and one account’s decisions never reach another account’s posts.
  • Changing or listing them. Tell the agent the new voice for that account; it saves the next version. Ask “what is saved for Maya’s account?” to list them. Account decisions are not listed in an agent’s Decisions section, because they belong to the account rather than to that agent.
To build a voice from what an account already posted, ask the agent to read the account’s recent posts and suggest a voice and topics; it saves them for the account once you approve them.

Seeing and retiring decisions yourself

Open an agent’s details. If you own the agent and it has saved decisions, a Decisions section lists each one with its version, the day it was made, what it is for (In one conversation, For this agent, For all your agents, For one Workflow) and where it came from: the Task, Page or memory it was taken from, or the room or DM you made it in (for example From Inbound), linked to your message there, as long as you can still open that conversation. Select Retire, then Retire decision in the confirmation, to take one out of use; Cancel or Escape keeps it. A retired decision can’t be brought back from this section: tell the agent the decision again. To change a decision, tell the agent the new one. Nobody else sees this section. When an agent changes owner, what its previous owner taught it stays with them: the new owner’s conversations with the agent don’t use it, and nobody else’s do either. What the agent learned in a conversation keeps the rule above, and what the new owner teaches it from then on is theirs.
In a room, an agent can still act on its owner’s accounts when someone asks, with approval as usual, and what it finds or does is posted in the room for everyone in it. Ask in your DM for anything you want to keep to yourself.

Records, Lists and Workflows in a room

An agent opens Records with its owner’s access, but in a room it reads only the Objects and their Records, Lists, notes and Workflows (with their runs) that everyone in the room can open. For example, if Yann’s agent Chief is in a room with Alfie, Chief doesn’t read an Object only Yann can open, or a List Yann hasn’t shared with Alfie, even when Alfie asks. Being a workspace owner or admin doesn’t change this: what the owner can open as an admin still has to be open to everyone in the room.
  • When a list or search leaves something out, the agent says some Records, Lists or Workflows were left out because not everyone in the room can see them.
  • When you ask for something not everyone in the room can open, the agent says so and suggests asking in the owner’s DM with it. It doesn’t name what it couldn’t open.
  • When you ask for an Object or Workflow by its name, the agent looks only among the ones everyone in the room can open. If none of them has that name, it lists those instead and says some were left out, without saying whether one you can’t all open has that name.
  • In a room open to the whole workspace, it reads only what every member can open, and nothing a guest in the room can’t open.
  • In a room shared through a public link, it reads no Records at all.
In a DM, the agent reads the Records you can open. With your own agent, that’s everything you can open, as before.

Existing work an agent is shown before it answers

Before an agent starts on a request, Connie looks for work that already exists in your workspace and that the request names: Tasks, Pages, Lists, Workflows, Objects and routines whose titles share words with what you asked. The agent gets a short list of them (at most eight, each with its title, status or description, and the day it last changed), so it can open the ones that matter instead of starting again. For example, ask “which website builder should we use?” when a Task says the site was published on Framer last week, and the agent sees that Task before it recommends anything.
  • The list holds titles only. The agent still opens an item to read it, with the same access rules as any other read.
  • Nothing is listed when no title matches your request. This works in your DMs with agents, in rooms, in routines and in work an agent does on a Task.
  • In your own DM, the list holds what you can open. In a room with other people, only what every member of the workspace can open, plus what belongs to that room. Another person’s agent, in your DM with it, is shown only what every member can open.
  • Nothing is listed in a room a guest reads or in a Slack thread.
  • Decisions and preferences you asked an agent to keep come from its memory, as described above, not from this list.

Purpose and reusable guidance

Give or change your agent’s purpose by talking to it. Its owner can update its internal Mission through conversation; Mission is not displayed or edited in the profile. There is no separate Instructions editor. Keep reusable facts and preferences in memories, and reusable procedures in skills. These keep their existing ownership and audience rules: a request in a shared room does not grant access to private memory or change an agent you do not own. You can ask an agent to read its saved Mission and standing rules and explain who may change them. They live in its saved Mission and versioned configuration, separately from its memory. Only its non-guest owner may change a rooms agent’s purpose or rules; room participation or a workspace admin role grants no ownership controls. Reading rules grants no permission to change them. An attempted change must return a saved result or a clear refusal; the agent confirms only a successful save.

What every agent knows about your business

Your workspace keeps one business profile: your company’s name, its website, and what it is, what it sells and to whom. Every agent in the workspace reads it, including agents you create later and agents your teammates own, so you say it once.
  • Where it comes from. The website or business description you give your Chief of Staff during onboarding is saved as the profile.
  • How to change it. Tell any agent, in its DM or a room: “We’re Northwind, an accounting firm for small restaurants, northwind.example.” It saves the profile for the whole workspace and every agent uses it from its next reply. Ask any agent “What do you have saved about our business?” to read it back.
  • Who can change it. A workspace owner or admin. A member’s agent reads the profile and tells them an owner or admin has to change it. Work an agent starts by itself (a routine, a trigger, a scheduled check, a recurring task, a watched Slack channel) never changes it.
  • What it is not. It is not an agent’s Mission. A Mission says what one agent’s job is; the profile says what company every agent works for. Correcting the profile does not rewrite any Mission, and there is no screen to edit either: you talk to an agent. It is also not where your decisions go: your ideal customer, pricing, approved copy or a booking link belong in the playbook, where each decision keeps its versions.
Earlier versions of the profile are kept. It is visible to every agent in the workspace, so don’t put anything in it that only some teammates should know.

Your personal memory

Your personal memory, the facts and preferences in Settings → Memory, is used only in conversations that only you read: your DMs with agents, and a room where you are the only person. It is never used in a room with other people, whoever’s agent answers there. See Memory to see, edit or turn it off. An agent keeps something you tell it to remember only in a conversation with one of your own agents that only you read. In a DM with a teammate’s agent, or in a room with other people, Memory is off: the agent doesn’t keep what you say there and won’t say it noted it. To keep something, tell it to one of your own agents in your DM with it, or add it in Settings → Memory. A memory kept from your message is a copy in your Memory, not a link to the message: editing or deleting the message later doesn’t change or remove it, and the agent goes on using it in conversations only you read. To change it, tell the agent the new version (it replaces the old one), ask it to forget it, or edit or delete it in Settings → Memory. A message you delete is no longer learned from: what Connie picks out of your messages in the background leaves deleted messages out, and an agent doesn’t learn from a message or an answer deleted in a room. When you Edit and resend a message in your DM with an agent, what was learned from that message and everything after it is set aside: memories picked up from them since are archived (one you edited yourself stays), and your earlier messages are read again, so only what still stands is learned. What the agent learned for itself from the removed work is set aside too.

Adding people shares a room’s history

Everyone in a room can read its whole history, including people added later. Adding someone shares what was already said in that room, and nothing from any other room or DM. The files, Records, Lists and other artifacts linked in it keep their own access: someone who can’t open one still can’t. Adding an agent to a private room does not add a new human reader or stop another agent’s answer. Work continues when the same people can read the room and its sources. A membership change alone does not cancel a task or undo actions it has already performed. If someone joins while an agent is still working on an answer that uses something they can’t read, the agent stops writing it in the room and posts Who can read this conversation changed while I was working, so I did not post that answer. Ask again and I will answer from what everyone here can see. instead. The same happens when the answer uses another room the agent read and someone here was removed from that room, or it was deleted, while the agent worked. From then on in that answer the agent also creates or changes nothing there, such as a Page, an artifact or a post: it is told “This uses something read earlier that not everyone in this room can read now, so nothing was written, created or sent here.” What the agent read earlier for fewer people isn’t brought back into the room later either. This also covers files from a code run. A code run in that answer doesn’t publish its files, such as a report or a CSV, to the room, and the agent gets the same message. An App edit also checks access before saving its changes, including edits that return no file. If you ask an agent in a room to bring in another room or conversation that not everyone here can read, such as its messages, people or tasks, the agent doesn’t bring it in. Its room tool answers: “Not everyone who reads this conversation can read that, so it is not brought here. Ask in a private conversation, or share one message with room-share-message.” The agent may pass that answer on to you. To use that content, ask in a conversation where everyone can read it, or share the one message you need as described below. This notice means the answer was withheld, not that the work was cancelled. Work may already have completed, and subsequent steps still check current access. Check Tasks and any changes already made before asking to repeat the action. Use Stop to stop work that is still running. Older work without a recorded reader list, and work using restricted context in a room open to the workspace, can still withhold an answer after an access change when safe sharing cannot be established. Adding people from an agent’s DM never shares the DM: it starts a new, empty room. See Rooms and people.

Share something on purpose

To bring something from a private conversation into a room, post it in the room yourself, for example paste the summary or the figures you want to share. Once posted, it is part of the room’s history for everyone in it, and only what you posted is shared. You can also ask an agent, in the conversation where it was said, to share one message or an exact passage of it into a room you can post in. It’s posted in that room as a quote from you. Nobody in the room gets access to the conversation it came from. Agents there see it as a passage you shared from another conversation, so they say it came from you, not that someone in the room wrote it; they don’t learn where it came from, who wrote it there, or anything else that was said there.

Good to know

  • Who manages an agent doesn’t decide what it shares. An agent’s owner change what it may do; the owner doesn’t get to read other people’s DMs with it.
  • Handing an agent work doesn’t get around this. When one agent asks another (or a step sends an agent a message), the agent asked recalls and reads only what everyone who reads its answer could already read: the person who asked and the conversation the request came from (and each one before it, when the work was handed on more than once), and from a room, everyone in that room. Asked from room C, Chief uses nothing from A or B, doesn’t open its owner’s private accounts, chats or memory unless everyone in C could, and when it looks back over its own earlier work it finds only what it did in conversations everyone in C can read, saying when some was left out. When it reads another agent’s work or memory, or names other agents, the same rule applies, and it does so as the person who asked, never as its own owner: it names, reads and changes only the agents that person may. The same holds after an approval in that work, and when it is asked from outside a room, for example from Slack or a scheduled chat. If, while an approval in that work waits, its conversation is shared with someone who can’t read the conversation the request came from, deciding the approval says “Who can read this conversation changed since this work was handed on, and not all of them can read where it was asked for, so it can’t go on here. Nothing ran. Stop sharing the conversation with them, or ask again.” and nothing runs; when only the person who asked reads the conversation the request came from, it goes on. And if, since that work began, someone may have come to read the conversation the request came from (someone joined the room, for example, the chat was shared or opened up, or a guest joined the workspace while the chat is open to it), deciding the approval while anyone but the person who asked reads the agent’s conversation says “Who can read where this work was asked for may have changed since it was handed on, so what it found can go on only in a conversation that just the person who asked reads. Nothing ran. Stop sharing the conversation, or ask again.” and nothing runs. Asked by a workspace owner or admin, it uses nothing only its owner reads. Its answer is in a conversation only the person who asked can open. When the request names one of the agent’s earlier conversations instead, it answers there only if that conversation is read by the same people as the one the request came from (from a conversation only the person who asked reads, any conversation they can open), and everyone who can open it counts among its readers; otherwise nothing is started there, and the agent can be asked again in a new conversation. A task in a room also goes on in a conversation one of its own earlier hand-offs used, even when fewer people read it than the room, while no one new can read the room since that hand-off and that conversation holds only that task’s work. Your agents don’t hand work to an old agent that didn’t become an agent here (see Old agents in Agents and permissions): everyone it’s shared with can read its conversations and activity, who may not all be able to read where the request came from, so nothing is started for it; hand the work to one of your agents instead.
  • An agent’s activity and memory follow the same rule. Wherever you read what an agent did, including through the API, you see only the work it did in conversations you can read now. Only its owner reads or changes its memory, and only the parts learned where the owner can read; workspace owners and admins included, nobody sees anything of a teammate’s starter agent’s work. When Connie reads them for you, for example in a scheduled chat or in Slack, or another agent does, it brings only what everyone who reads its answer could read: in a chat shared with other people, only the work done in conversations all of them can read, and in a Slack channel, nothing that isn’t public.
  • Owners’ and admins’ information, such as the workspace’s credit usage, its subscription or another member’s account settings, is read in a room only when everyone in it is a workspace owner or admin. Otherwise the agent answers with what every member may see, or says the rest is only for owners and admins.
  • An agent doesn’t say what it can’t share. Asked about other conversations, in a room or in a DM with a teammate’s agent, it says it only knows what is in this conversation; it doesn’t confirm that other rooms or conversations exist, or who else talks with it.
  • Previews and notifications follow the same rule: a room’s preview in the sidebar and its notifications only ever show that room’s own messages.