Long conversations, no lost context

Choose how the platform handles long conversations: Off, Summarise & Start New, or Auto-Continue.

Off: you manage conversation length manually. Summarise & Start New: when the context limit approaches, the platform generates a summary and shows it to you before creating a new conversation — you can edit the summary before continuing. Auto-Continue: the platform generates the summary and creates the continuation automatically, with a 5-minute undo window. Settings apply at five levels: platform default → tenant default → hierarchy node → user preference → per-conversation override.

A thin progress bar in the conversation header shows how much of the model's context window you have used.

The bar is visible in every conversation regardless of your continuation setting. Colour transitions: green (0–49%), yellow (50–74%), orange (75–89%), red (90%+). Hovering shows the exact percentage and your current continuation setting. At 80% utilisation, an inline system message appears in the conversation thread. At 90%, a second system message confirms that a continuation is close.

When continuation triggers, an AI-generated summary captures the key points of the conversation for the next session.

Summary generation uses the cheapest classification-eligible model — not your selected model — and is attributed to the same cost code as the conversation. In Summarise & Start New mode, you see an editable text area pre-populated with the summary before you confirm. The summary carries forward to the new conversation as a system message. Continuation chains are linked via conversation IDs so you can navigate the full history.

Each continuation creates a new conversation linked to the original. You can navigate the full chain from the conversation header.

Continuation naming convention: [Original Title] (cont. 1), (cont. 2), and so on. The chain is atomic for GDPR erasure — when you request erasure of a conversation, all continuations in the chain are erased together. You cannot erase only part of a chain. Branches off a continued conversation are tracked independently from the main chain.

Agent workflows use Auto-Continue automatically. There is no Summarise & Start New option for agent runs.

Long agent tasks often exceed context limits mid-run. Auto-Continue is the only safe option for agents: the platform cannot pause to ask a human. The continuation is attributed to the agent's cost code and logged to the same audit trail as all other agent actions. Agent continuation chains are separate from user conversation chains.