Semantic Memory

What is semantic memory?

In psychology, semantic memory is our memory for facts: you know water freezes and Paris is in France, without remembering the day you learned it. In AI, semantic memory is the stock of factual knowledge and concepts an agent keeps — definitions, rules, common sense, domain knowledge.

It pairs up with episodic memory

Semantic memory: "what is"
Stable knowledge detached from any scene. "Expenses need a receipt" holds true everywhere.
Episodic memory: "what happened"
Experience with time and context — "last week you did my expense report and it got bounced for a missing receipt". One handles knowledge, the other handles experience.

Where it comes from and how it's stored

Many sources
Hand-written rules, or knowledge extracted from documents and organized by a model.
Usually a vector database
Knowledge is turned into vectors and retrieved by semantic similarity. Ask "how do I get a refund" and it surfaces the "refund policy", even if the wording differs.

What problem it solves

Without semantic memory, an agent can "forget common sense" from one answer to the next, or contradict itself. With it, the agent has a stable knowledge base — it follows rules consistently and cites the right concepts instead of winging it every time.

Bottom line: semantic memory is the AI's "knowledge base" — it remembers "what is", while episodic memory remembers "what happened". Together they make a full memory.