Agent conversation state contains the active interaction history that helps a model understand what has already happened. It may include user messages, assistant replies, tool outputs, plans, errors, and temporary conclusions that have not been written into durable project files.
This state may disappear when a provider session ends, a context window is replaced, or a different model starts a fresh session. Durable handoff notes and repository files preserve important information without assuming that the next session can recover the previous conversation.
ELI5
Agent conversation state is the working memory of one AI session. It contains the recent discussion and tool results that let the agent continue from the last step without asking for everything again.
For example, a fresh provider session may still see the same code files but forget why one option was rejected earlier. Writing that decision into a handoff file lets the next session continue without relying on vanished chat history.
Is conversation state the same as repository context?
No. Conversation state is session-specific interaction history, while repository context is stored in project files and related durable state that another session can inspect.
Why can conversation state disappear?
A session can end, a provider can start a new context, or context limits can remove earlier material. Important decisions should therefore be preserved in a durable form.
