Why AI Agents Need Both Episodic and Semantic Memory to Avoid Costly Mistakes

AI agents used in scheduled automations can fail in two distinct ways: by knowing the rules but forgetting past events, or by recalling every incident without ever extracting useful lessons from them. This distinction maps to cognitive science concepts formalized by Tulving in 1972 and later applied to AI agents in the 2023 CoALA framework, which separates episodic memory (timestamped event logs) from semantic memory (general facts and rules). A practical example is an agent that correctly applies a refund policy but re-issues a refund already processed in a prior run, simply because it has no record of that earlier action. Most DIY agent setups rely on only one memory type — either a vector database of facts or a chat-history log — leaving them vulnerable to the failure mode of whichever type is missing. Effective scheduled automation requires both memory systems working together, along with procedural memory for learned routines, to ensure agents can reason accurately across runs.