Skip to main content
Your agent is stateful in a few ways: the configuration that defines it, the conversations it holds, the per-user progress it tracks through a workflow, and the memory it keeps per user. This page describes what’s kept and where you manage it.

​
Conversations per channel

Each channel keeps its own conversations, and every conversation is retained in full. The same user returning on the same channel — the same WhatsApp sender, the same email address — resumes their existing conversation, even after a restart or update. A brand-new dashboard or web chat starts a fresh conversation each time. Every conversation from every entry point — web, widget, API, A2A, email, schedules, webhooks, IM, and more — shows up in the owner Inbox in one place, and its full transcript and turn-by-turn record are viewable together (see Observability and debugging). The same person on two different channels — say Telegram and email — is treated as two separate identities, with separate conversations and separate memory. They’re unified only for viewing, in the Inbox and Contacts pages. The per-channel identity rules are documented in User memory.

​
Per-user workflow progress

If your agent uses stages and goals, each user’s progress is tracked separately, alongside their memory: You author the workflow — stages, goals, and how a user moves between stages — on the Purpose page, and configure goal checks on the Goals page. A user only advances past a stage’s required goals once those goals are met.

​
Configuration history

Every time you save a configuration section, the previous version is kept, so you can roll back. The most recent versions of each section are retained, covering every save from the agent editor, the Designer Chat, template application, and restores themselves.
  • Identical saves don’t add clutter — saving unchanged content is skipped.
  • Restores are safe — a restore is checked against the current section format before it’s applied, and the state before the restore is itself saved, so a restore is undoable.
  • Clones start fresh — cloning an agent copies its configuration but not its history.
You access and restore history from the agent editor.

​
What you can rely on

  • Your saves are never left half-written. A restart mid-save leaves either the old version or the new one — never a corrupted state.
  • Configuration edits take effect immediately, with no restart.
  • Approved actions run at most once, even across a retry — see Tool execution and reliability.
A completed external action — a sent email, an IM message, a write through a connector — has no undo. If the surrounding turn later fails, that action still stands. Design flows so that irreversible actions happen as late in a turn as possible, and see Tool execution and reliability for the consent gates and guards that control when they fire at all.