What is Semantic memory?

Semantic memory is the knowledge half of an agent's long-term memory: durable facts about the world it operates in, decoupled from the specific moment they were acquired. Where episodic memory records that something happened at a particular time, semantic memory records what is true in general, this service authenticates with a bearer token, the team prefers tabs over spaces, the staging database is a branch of production. The term comes from cognitive science, where semantic memory is your general knowledge (Paris is a capital city) as opposed to episodic memory of personal experiences. For an AI agent, semantic memory is what makes it competent on a domain over time: it accumulates conventions, preferences, decisions, and the rationale behind them, then surfaces the relevant ones into the context window when a task needs them. These facts are often distilled from many past episodes, the agent notices a pattern across events and stores it as a general rule. A team-shared semantic memory is especially powerful, since a fact one agent learns becomes knowledge every teammate's agent can rely on without rediscovering it. Glen captures durable facts as atomic observations and retrieves the relevant ones on demand, which is semantic memory delivered through an MCP tool.