What is a Semantic Memory (AI)?

This mirrors the cognitive-science distinction between semantic memory (general facts — "Paris is the capital of France") and episodic memory (specific remembered events — "we discussed the Paris trip last Tuesday"). In AI agents, semantic memory is typically implemented as a structured or vector-embedded knowledge store that persists facts about a user, company, or domain across sessions, distinct from the short-term memory of the current conversation. The practical benefit is continuity without re-explaining: an agent with semantic memory of a company's product catalog, org structure, or stated preferences doesn't need those facts restated every session — it retrieves them when relevant via similarity or structured lookup. This is usually built on top of a vector database for semantic search over embedded facts, sometimes paired with a structured graph for explicit relationships. Semantic memory needs active curation: facts change (a person's role, a product's price), and a memory system without update and conflict-resolution logic will confidently retrieve stale information, which is often worse than having no memory at all.