When should I clear the chat?

When the topic changes. The tell-tale sign is writing "forget the previous thing" or "now about something else". Press "Clear chat" in the chat menu or send /clear. The messages are deleted and the agent starts from a clean slate, keeping what it has already remembered. Right after clearing you can bring them back with "Undo".

What does the “Compacting context” step mean?

The conversation grew too long, so the agent condenses its earlier part into a summary on its own and then carries on answering. Your messages and conclusions are kept, the details of old exports are not. The feed stays as it was. It is a signal to clear the chat.

Can I compact the context by hand?

Yes, with the /compact command while the agent is not answering. After a space you can name what the summary should keep in detail: /compact decisions on the 1C integration.

What if the context could not be compacted?

Clear the chat from the chat menu or with /clear. If the message itself is too large, send a shorter one and attach a big export as a file. There is no point in repeating /compact: the agent has already tried compacting the context the same way.

Chats and context

A chat is a desk, not an archive. The agent re-reads everything in it at every step: the longer the conversation, the more attention goes to old material and the higher the chance yesterday's report bleeds into today's.

As many chats with an agent as you like

You can run as many chats with one agent as you like: one chat per task works best. Earlier chats stay in the list on the left.

Every agent has a main chat: it appears in the list on the left as soon as the agent is created and is labeled with the agent's name. Messages from other agents arrive there. For an agent a colleague shared with you, Start a chat in Settings → Agents opens its main chat.
New topic: New chat, and the task starts from a clean slate. Or Clear chat in the chat menu (the "⋯" button in the header or the row menu in the left panel), or the /clear command.
Clearing means deleting, as in any messenger: all messages leave the feed and the chat stays in the list, empty. Changed your mind right away? Press Undo in the notice that appears.
The agent saves what mattered in the cleared chat to memory, so you do not repeat facts.
The sign it is time: you write "forget the previous thing", "now about something else", "let's start over".
A shared chat with other people has no clear: everyone in it reads the same feed, and clearing it would erase it for all of them.

Rename a chat

Open the chat's row menu in the list on the left and choose Rename chat. The name appears as a label next to the agent's name; the agent's own name stays the same. Leave the field empty and the chat is labeled by its topic again. A group chat without a name is called by its members.

If the conversation did grow long

The agent compacts the context on its own, there is nothing to press. While it does, the turn's status line reads Compacting context, and then the answer continues. Your messages and the conclusions reached are kept in the summary; the details of old exports are not. The feed stays as it was: only what the agent re-reads is compacted.

Signs of an overloaded chat

Answers get slower, the agent mixes up periods and files, or returns to a task you closed earlier. Clearing the chat and giving a short summary in the first message is faster than untangling the long one.

Compacting the context by hand

The /compact command compacts the context right away, without waiting for the conversation to reach the threshold. After a space you can name what the summary should keep in detail: /compact decisions on the 1C integration. While the agent is answering the command will not run: repeat it once the answer is finished.

If the context could not be compacted

Rarely, the conversation does not fit the model's window even after compacting, usually because of a very large message or paste in the current request. The turn then stops with the error “The chat no longer fits the model's window”. What to do:

clear the chat from the chat menu or with /clear: the agent starts from a clean slate and keeps what it has remembered;
if the message itself is too large, send a shorter one and attach a big export as a file: a file goes to the workspace, not into the conversation.

There is no point in repeating /compact: the agent has already tried compacting the context the same way.

What a long chat costs

Some models have a price step: a request longer than, say, 272 thousand tokens is billed at double the rate. The agent compacts the context before that step, so a conversation usually never reaches it.

A long chat also uses up the weekly limit faster: the agent re-reads the whole conversation at every step, and every new message is billed together with everything it re-reads.

What to say after a clear

Give the agent the minimum the task needs:

the file itself: attach it again, that is more reliable than "the file above";
one paragraph of conclusions from the previous conversation, if they matter;
anything that should always apply: write it to memory, then you never repeat it again.

Asking about part of an answer

Select a fragment of the agent's answer and press Ask. The fragment appears above the input as a quote, and your question goes out as a reply to exactly that: "where does this number come from?", "rewrite this paragraph shorter". No need to copy the text or explain which part you mean.

Incognito mode

Incognito mode is for a one-off sensitive question: the chat is not saved to history and nothing is written to memory. It is not meant for everyday work, since the agent cannot build on that chat later.