Episodic memory records specific events, semantic memory stores general facts, and procedural memory holds how to do things; AI systems borrow all three.
The three-way split comes from cognitive psychology. Episodic: 'on Tuesday we decided to drop the Redis option'. Semantic: 'the project uses Postgres'. Procedural: 'to deploy, run the release script then tag'. In AI memory systems, episodic entries look like dated checkpoints and logs. Semantic entries look like a fact list or knowledge graph. Procedural entries look like instructions, playbooks or a project brief that tells an agent how the team works.
The distinction is useful for design. Episodic memory grows without bound and needs compaction. Semantic memory should be small and current, and old facts should be replaced, not appended. Procedural memory changes rarely and should be loaded at session start. Systems that store everything as one undifferentiated log make all three harder to use. Systems that keep typed entries can load procedure first, facts second, and search episodes only when asked.