Semantic Memory

Semantic memory holds structured facts the agent knows about the world: "Alice prefers email over phone," "the company's refund policy allows returns within 30 days," "the customer's primary product is Plan A." It differs from episodic memory in two ways: it is not tied to a specific event (semantic facts are timeless), and it is structured rather than narrative (key-value or graph rather than transcript). Storage is typically a JSON document store, a vector database for unstructured facts, or a knowledge graph for highly-connected facts.

The cleanest production pattern: episodic memory captures what happened, a separate process distills it into semantic memory facts. After the call where Alice complained about the restocking fee, an extraction step writes "Alice dislikes restocking fees" to her semantic memory profile. Future agents read that profile before responding. This separation keeps episodic memory full of raw history (good for audit) and semantic memory full of distilled knowledge (good for reasoning), without polluting either.

Use semantic memory for stable user preferences, account state, and entity attributes. Pair with episodic memory for the full picture.

• LangGraph: Memory concepts
• Letta: Memory architecture

Related Terms