What the system remembers between turns. Some of it is the conversation, some is data gathered along the way, some is where the agent had got to in a longer job. Being explicit about what persists, where it lives and when it is cleared prevents a whole category of confusing behaviour.
Technical view
Distinguish conversation history, working memory such as intermediate artefacts, and durable memory across sessions. Storage sits outside the model and is reconstructed into context each turn, so context assembly policy is the real definition of what the system knows. Concerns include size growth requiring compaction, consistency under concurrent requests, expiry, and the fact that session stores frequently contain personal data and therefore fall within retention and deletion obligations.
Related
Part of AI Terminology Index, section: Agents and Tools.