AI agents require four memory types to overcome LLM statelessness

AI agents built on large language models are inherently stateless, forgetting all information between each API call. To become useful, such agents need four integrated memory systems: in-context, external, episodic, and semantic. Each system handles a different timescale and retrieval pattern, storing information from conversation threads, user preferences, past mistakes, and domain knowledge. Together, they create the illusion of a persistent, learning assistant that can remember and adapt across sessions. The concept is illustrated using Mailmind, a fictional AI email assistant.