Semantic memory
The agent's store of timeless facts — "the user is a VP of Sales at Acme," "the company uses Snowflake" — distinct from events (episodic) or skills (procedural).
Semantic memory holds facts that are true across time. Where episodic memory stores "yesterday the user said they prefer dark mode," semantic memory stores "the user prefers dark mode." The latter survives until something contradicts it.
Production agents extract semantic facts from each session — usually with an LLM extraction prompt — and merge them into a key-value or graph store. Conflict resolution is the hard part: which version of a fact wins when they contradict?
Semantic memory is the layer that makes an agent feel like a colleague rather than a stranger. Combined with episodic memory, it is the foundation of any "memory-enabled" assistant.
Frequently asked
How do I extract semantic facts from a conversation?+
A small LLM prompt at session end: "From this conversation, list facts about the user that will be true tomorrow." Filter for stability — preferences, role, employer — and skip transient state.