A language model has no memory of its own. Each request starts from nothing, and anything that looks like remembering is really information that was put back into the prompt. Conversation state is the history of messages and results you send with each turn, and memory is what you store outside the model so it can be recalled later.
The lessons separate short-term memory, which lives in the context window and is lost when the session ends, from long-term memory, which is saved and retrieved across sessions. You will also meet the common types of long-term memory (facts, past events and learned instructions), and the decisions that come with them: what to save, when to write it, and how to keep a growing history from filling the context window.
Watch the lessons in order, then continue on to Tools & Agent Context.
Google Cloud Tech's short explainer on why agents need memory, and how short-term, long-term and working memory differ, with an example of how to implement each.
Video 4minLangChain's CEO on giving agents long-term memory: semantic (facts), episodic (past interactions) and procedural (instructions) memory, hot-path vs background memory updates, and a worked email-assistant example.
Video 1h 31minA conversation with an OpenAI solutions architect on managing conversation state in tool-heavy agents: trimming and summarizing history, short- and long-term memory, and a live demo comparing an agent with and without context management.
Video 57min