Skip to content
Strategy Agent Docs

Conversations you can come back to

1 November 2026

Conversations now keep what was said in them:

  • The sidebar of the Web App lists your Conversations in the current Tenant, the one you worked in last on top. Each gets a short title from the agent after its first answer. Rename a Conversation with the pencil next to it, and search your Conversations by title. See Work in Conversations.
  • Delete a Conversation with the bin next to it, after confirming. It leaves your list at once, can’t be restored, and everything in it is erased for good after 20 days. The Audit Log records the deletion with the Conversation’s ID only. See Delete a Conversation.
  • Open a Conversation again, or reload its page, and it shows every earlier Turn as it was. Long Conversations load their newest Turns first and older ones as you scroll up. See Work in Conversations.
  • The agent goes on working when you close the tab or lose your connection. Come back to the Conversation to see the answer so far and follow the rest as it comes in, or the whole answer once the agent has finished. The sidebar marks the Conversations the agent still works in with Turn running. See Close the tab while the agent works.
  • Long Conversations keep working. When a Conversation grows long, the agent works from a summary of its oldest part, while you still see every Turn. See Stay in one Conversation as long as you like.
  • The message field stays at the bottom of the screen, and only the Conversation scrolls. It follows an answer as the agent writes it, until you scroll up; Scroll to the newest message takes you back. On a phone, your Conversations open from Show your Conversations in the title bar, so the Conversation gets the whole screen. See Write and read in a Conversation.
  • The agent’s answers are formatted, with tables, lists, headings and bold text. Links open in a new tab, and each code block has Copy code. See Answers with tables, lists and code.
  • A new Conversation offers a few Suggestions of what the agent can do, and after each answer a few Suggestions follow from it. Select one, and it is sent at once. See Start from a Suggestion.
  • Select a passage of an answer to Quote it into your next message, so the agent knows which part you mean, or to Copy it. Your message shows the quote above its text. See Quote part of an answer.
  • A Conversation that was deleted, never existed or belongs to someone else shows the same page, This Conversation isn’t available, so nobody learns whether someone else’s Conversation exists.
  • The agent can ask you a question when it needs your answer to go on: select one of the answers it offers or write your own, and it goes on with the same Turn. Until you answer, the message field is blocked, also after a reload. See When the agent asks you something.
  • Before a tool that changes something runs, the agent asks for your Approval: accept it, edit its inputs first, or reject it, and the agent goes on with the same Turn. Until you decide, the message field is blocked, also after a reload. The tools the agent has today only read, so Approvals show once tools that change things arrive. See Approve a tool before it runs.
  • The agent’s reasoning and the tools it uses show in its answer as steps, in one block that is open while the agent works and closes once the answer goes on. Reasoning shows what the agent thought, Thinking… while it still thinks, and each tool shows as a line you can expand to see its inputs and result. See See what the agent did.
  • A Turn that fails keeps what the agent wrote so far and tells you whether the agent ran into a problem or the connection was lost. When the agent ran into a problem, the message shows inside the answer that failed, with Continue next to it. Select Continue, and the agent picks up where the Turn ended; a line marked Continued shows where it began. See When a Turn fails.
  • Stop a Turn with Stop, which takes the place of Send while the agent works. The agent stops within a few seconds, keeps what it wrote so far and marks a tool it was running Cancelled; Continue picks the Turn up again. Deleting a Conversation stops its running Turn too. See Stop a Turn.
  • The agent remembers what you tell it to: say Remember that I want numbers in millions, and it knows that in every Conversation of yours in the Tenant. A line among its answer’s steps confirms what it saved, with Undo. Tell it to forget something the same way; Undo brings it back as something you told it. Nobody else sees what it remembers. See What the agent remembers.
  • The Memory page, in the account menu, lists everything the agent remembers about you: edit or delete each memory, or delete them all. Switch Memory off there, and the agent neither uses nor saves memories, while what it remembers is kept and you can still tell it to forget something. See See and change everything the agent remembers and Switch Memory off.
  • The agent also learns on its own: about half an hour after a Conversation went quiet, it remembers what’s worth keeping from your messages and its answers, never from its reasoning or its tools’ results. What it learned shows on the Memory page marked New, with a link to the Conversation it came from. It never changes what you told it to remember or corrected yourself, and it learns nothing while Memory is off. See The agent also learns on its own.
  • The agent works on one Turn of a Conversation at a time. A message sent while another Turn of the same Conversation still runs, for example in a second tab, is turned down with a message that says so.
  • Rate any Turn that has ended 👍 or 👎, under its last answer. After a 👎, tell the agent what went wrong with a few reasons and a comment, and it follows up in the Conversation to put it right. Change or remove your Feedback at any time. A Turn you rate becomes readable, the agent’s reasoning included, for users of your Tenant with the Right Read rated Turns and for our Super-admins with the matching Platform Right, until you remove your Feedback. The reasoning can contain data the agent read, such as rows from a Data Source or the contents of a web page. See Rate the agent’s answers and Privacy.
  • The agent can search the web when your question needs current public information, read a page you give it, and open pages that need a browser in the AgentCore Browser. It shows what it searched for and the sources it read, which open in a new tab. It reads only public pages, never signs in anywhere and treats what it reads as information, never as instructions. Web search runs on Amazon’s own web index in Ireland. See Let the agent use the web and Data residency.
  • Admins read the Turns users of their Tenant rated on the new Rated Turns page of the Admin App: filter them by Feedback, reason and day, and open one to read the question, the answer and its steps, the agent’s reasoning and the tools it used, without the rest of its Conversation. The page needs the Right Read rated Turns, and the Audit Log records every opening. See Rated Turns.
  • Authorised Scavenger AI staff read rated Turns of every Tenant to improve the agent, under the same rules: one Turn at a time, only while its Feedback stands. Only the staff members granted the right to read rated Turns can, also on your Tenant’s own Rated Turns page. Each time they open one of your Tenant’s, its Audit Log records it, marked Scavenger AI staff. See Our staff.