Semantic memory is a type of AI agent memory that stores general facts and knowledge independent of when or where they were learned. Borrowed from cognitive science, it contrasts with episodic memory, which records specific events in sequence. Semantic memory is the agent’s store of what is true, not what happened.
For an AI system, semantic memory holds durable knowledge: facts about the world, a user’s stable preferences, domain information, and learned concepts. These are retrieved when relevant to the current task, regardless of when they entered the store. A vector database is well suited to this, embedding facts so they can be found by meaning whenever a related question arises.
In practice, capable agents combine semantic and episodic memory. Semantic memory supplies the general knowledge needed to reason, while episodic memory supplies the specific history of interactions. Together, backed by vector retrieval, they let an agent both know things and remember what it has done, behaving consistently and drawing on accumulated understanding across many sessions.