Explain the four types of memory in production AI agent...

Four distinct kinds, and conflating them is where most memory designs go wrong:

• Working memory — the current conversation and task state, living in the context window. Fast and free to access, but gone the moment the session ends. Think RAM.
• Episodic memory — records of specific past interactions: “last time this user asked about pricing, they wanted the enterprise tier.” Stored externally, typically in a vector database, and pulled in by semantic similarity to the current query.
• Semantic memory — general facts and domain knowledge the agent has accumulated, independent of any particular conversation. This is closer to a knowledge base than a chat log.
• Procedural memory — not facts but procedures: the sequence of actions that worked last time for a given kind of task. A support agent that has learned “for a delayed-order complaint: check tracking, check warehouse, offer reship or refund” is exercising procedural memory, not recalling a fact.

A production agent typically layers all four rather than picking one: working memory for the current task, episodic for personalization, semantic for domain grounding, and procedural for not re-deriving the same successful workflow from scratch every time.