Multi-Turn Conversation — AI Glossary

Reference details

Related terms

• Conversation History
• Message Threading
• Context Management

How it works

1. The endpoint is stateless, so the client resends the whole transcript on every call — the conversation exists in the request, not on the server.
2. Each turn is role-tagged and the template flattens them in order, which is why earlier constraints remain visible as tokens rather than as remembered state.
3. Input grows with every exchange, so cost per turn rises across a conversation even when the new message is short.
4. Once the transcript approaches the window, something must give: truncation drops the oldest turns, or summarisation compresses them and loses detail the user may still rely on.
5. Attention is not uniform over a long transcript, so a constraint set early can be followed less reliably than one restated recently — which is the failure single-turn evaluation cannot see.

Example: Why cost grows quadratically with the conversation

Each turn adds about 200 tokens. Turn one sends 200, turn two 400, turn ten 2,000 — and the total sent across ten turns is 200 x (1+2+…+10) = 200 x 55 = 11,000 tokens, not the 2,000 the final message suggests. Twenty turns is 200 x 210 = 42,000, so doubling the length nearly quadruples the spend. The sum grows with the square of the turn count, which is the arithmetic behind summarising history rather than resending it indefinitely.

Commonly confused with

Describes the interaction, not the payload. Conversation history is the mechanism that makes it possible, since chat endpoints keep no state between calls. Evaluating a model on single-turn benchmarks says little here: the failures specific to multi-turn are forgetting a constraint set several turns earlier and mishandling a topic shift, and neither is visible in a one-shot test.

When to use it

Reach for it when:

• Evaluating whether a model holds constraints stated several turns earlier.
• Designing history handling — truncation versus summarisation — before the window is hit.
• Budgeting a conversational feature, where cost per turn is not the cost of a turn.

Reach for something else when:

• Single-shot tasks such as classification or extraction, where no history exists.
• As a description of the payload — the mechanism is conversation history, not the turn count.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Multi-Turn Conversation — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path