Chat and Task Settings¶

Every chat has a set of switches that decide what the agent is given with each message and which platform features the chat has. A task carries the same switches: a chat started from the task copies them, and Update available copies them again (see Tasks and Update available).

The switches are in the chat's side panel under Settings, and in the task editor under its Settings tab. Some of them change the agent's behaviour more than the goal does, so choose them deliberately — especially for Live Apps.

Every CLI option also has a --no-… form, for example --no-chat-history. pt task create and pt task update accept them all.

History¶

On: with every message, the agent is given the last 50 messages of the chat as the conversation so far. That includes hidden messages — the task messages a Live App sends with pt.addMessage(…, { hidden: true }) and the agent's replies to them — and messages that another agent turn is still answering.

Off: the agent is given only the message it is answering. It remembers nothing from earlier messages, except what is kept elsewhere: the goal, the memo, memory, Chat DB entities and documents. A Deep1 agent can still read the chat summary (/chat/summary.md) and the memo.

Watch out:

• A Live App that sends several hidden tasks at once should switch History off. With History on, each agent turn sees the other tasks that are still waiting and may carry them out as well: in a test, four of six turns sent at the same second also performed the other five tasks. With History off every turn sees only its own task. Put everything a task needs into the message itself or into Chat DB.
• History is part of every turn's input, so switching it off also makes each turn cheaper — noticeably so in chats with long messages.
• A conversational assistant, where people refer back to what was said, needs History on.

Search in Chat Messages¶

On: for every message, PrimeThink searches the chat's older messages (those outside the last 50) for passages related to the new message and adds them to the agent's prompt. It also gives the agent the chat's automatic summary.

Off: no older messages are searched and the summary is not given to the agent, even when Auto generated summary keeps writing it.

Global Memory¶

On: the agent can save and recall long-term memories about the user, and relevant memories are added to its prompt. It needs the agent's Memory capability. See Memory.

Off: no memory tools and no memories in the prompt. Memories already saved are kept.

Documents and Collections¶

On: the agent is told which documents and collections are attached to the chat (a Deep1 agent sees them under /chat/documents, with an INDEX.md), and AutoRAG can be switched on.

Off: the agent is no longer shown the chat's documents and collections, and AutoRAG cannot run.

Watch out:

• This is not an access control. An agent with the Documents capability can still read and write the chat's documents with its tools, a Live App can still use its document functions (pt.getDocumentText, pt.documentUrl, uploads), and the API is unaffected. To keep an agent away from documents, remove the capability or the documents.
• A task published with pt task publish or pt live-app publish has it off unless you turn it on (see the warning above).

AutoRAG¶

Requires Documents and Collections.

On: for every message, PrimeThink searches the chat's documents whose status is Search and its indexed collections, and adds the best-matching passages to the agent's prompt before the agent starts. See Documents Attached to a Task for document statuses.

Off: the agent finds document content only when it decides to look, with its document tools.

Watch out:

• Every message pays for the search and for the larger prompt, and the agent starts a little later. Keep it for knowledge-base assistants whose answers come from the documents.
• In a Live App chat driven by structured hidden tasks, the added passages rarely help and add tokens to every task: leave it off unless the app's tasks answer questions from documents.

Auto generated summary¶

On: a background job writes a running summary of the conversation, roughly every 10 new messages or every hour.

Off: no new summary is written; the existing one is kept.

The summary reaches the agent only when Search in Chat Messages is on (a Deep1 agent can also read /chat/summary.md). With Search in Chat Messages off, the summary is written but not used.

Scheduled Tasks¶

On: the agent can create scheduled tasks from the conversation — for example "remind me every Monday at 9". It also needs the agent's Scheduled Tasks capability. Scheduled tasks are listed in the chat's Scheduled Tasks tab; see Scheduled Tasks.

Off: the agent cannot create new scheduled tasks.

Watch out:

• Switching it off does not stop existing scheduled tasks. They keep running; pause or delete them in the Scheduled Tasks tab.
• A task's own schedule is set up when a chat is started from it, whatever this switch says.
• A Live App whose agent schedules jobs (a nightly sync, a periodic sweep) needs this switch on in the chats that run the app — set it on the task.

Email¶

On: the chat gets its own email address, shown under the switch with a copy button. An email sent to it from the address of a group member who is in the chat arrives as that person's message, with attachments added to the chat's documents, and the agent's reply is emailed back in the same thread. A task can have an address too. See Email Integration.

Off: the address stops accepting email.

Make Public¶

On: the chat gets a public link. Anyone with the link can use the chat without logging in. Each visitor gets their own sub-chat and sees only their own conversation — never the main chat or other visitors' sub-chats. The sub-chat takes the main chat's goal, agent, settings, documents and collections. Switching it off disables the link. On a task, Public Chat makes the chats started from it public.

Watch out:

• The agent answers visitors with the chat owner's account. Whatever that agent can reach — its tools, the attached documents and collections, the owner's memory scope — anyone with the link can ask it about. Give a public chat an agent that has only the capabilities the public use needs, and do not attach private documents.
• Visitors are rate-limited, per visitor and per link.

Tasks and Update available¶

• A chat started from a task gets the switches of the task's Production version.
• Update available copies all of these switches, plus the agent and the goal, from the task's Production version into the chat — replacing any change made on the chat itself. Make the change on the task and publish a new version instead of changing it in each chat. See Updating Existing Task Chats.
• When an update exists and you can edit the chat, an Update available button appears in the chat's Settings panel and a reload icon marked with a red dot (tooltip Task update available) appears in the chat's top bar — on the chat's tab when tabs are showing. Either one opens the same dialog.
• The memo and the summary are never copied.

Recommended settings¶