Conversations

A conversation is the durable runtime record for an interaction with an agent. It contains transcript, context, active capability state, structured cards, artifacts, delivery events, and human participation.

Runtime lifecycle

1. Open a new conversation or resume an existing one.
2. Send a message or structured card interaction.
3. Receive an inline response or stream delivery events.
4. Reconcile history after reconnecting.
5. Continue until completion, escalation, handoff, or abandonment.

Human takeover does not create a separate conversation. Ownership and execution state change while the same transcript and audit history continue.

Agent working memory

An agent can retain concise facts, attempted steps, reported outcomes and what to do next within a conversation. Separate issues can have separate records, so returning to an earlier topic does not require repeating its details.

This memory is separate from capability datapoints and execution state. A memory update does not change an active capability's inputs, confirmations or phase. Capabilities continue to collect and validate their own inputs.

Goals can optionally assign typed datapoints to capture from participant messages. Each captured value retains its source evidence; corrections replace the current value for that issue without erasing its history. A rejected correction does not store the proposed value; the earlier value becomes unavailable until valid participant evidence captures it again. Capture happens alongside the agent's existing decision, without a separate extraction call. When captured details feed an action, the runtime validates and freezes the selected issue, goal, definition and context revision for that action.

For a direct copy, the agent can bind an input to a captured detail and leave its value for the runtime to supply. The runtime uses the selected issue's current, validated value. An explicitly supplied stale or reworded copy is rejected. A separately composed summary uses the action's ordinary approved inputs. Complete input validation and verified identity bindings still apply.

Agent instructions can describe what to remember and how to distinguish an offered suggestion from a customer-reported success. Updates retain supporting message or tool references. These references support inspection; they do not independently verify every interpretation. Current external facts still need appropriate tools, and knowledge answers still need current supporting sources.

Working memory is bounded and scoped to the conversation, agent and participant. It is not shared across conversations or exposed through the public runtime SDK.

Authorized Studio conversation traces show Captured details as Agent capture, with current values, source evidence, correction history and rejected updates. Capability values remain in their own invocation trace. This private audit view is separate from customer chat and public runtime events.

Applications should use the runtime SDK unless they need to implement the lower-level runtime contract.