TYPES OF AGENT MEMORY

EPISODIC MEMORY VERSUS SEMANTIC MEMORY Psychologists divide human long-term memory into types, and agent designers borrow two of them. Episodic memory is memory of specific past events, each tied to a time and a context, like diary entries. An example is "On March 3rd, the user rejected the blue color scheme and asked for green instead." It records what happened, when, and sometimes why. Semantic memory is memory of general facts, detached from the specific episode in which they were learned. An example is "The user's brand color is green." It is what you know, as opposed to what happened. Notice how the two are related. The semantic fact was extracted and generalized from the episode. Humans do this all the time. You may remember that Paris is the capital of France without remembering the day you learned it. There is a third type, sometimes added, called procedural memory, which is memory of how to do things, such as a learned recipe or workflow. In agents, this often takes the form of saved instructions or skills, which we explore in Chapter 48. Why does the distinction matter in practice? Because they are stored and used differently, and because mixing them up causes mistakes. Episodic memories are useful for questions like "What did we decide last time, and why?" They are time-stamped and specific. Semantic memories are useful for everyday decisions like "What format does this user like?" They are short, general, and quick to apply. A common design mistake is storing a one-off event as a permanent fact. Suppose a user once complained, in an angry moment, "I hate this long format." If the system stores "the user hates long formats" as a lasting fact, it will keep applying that to every future task, even though the user may have only been frustrated about that one report. Good systems keep the episode as an episode, and only promote it to a lasting semantic fact when there is enough evidence. That evidence might be the user stating the preference explicitly and generally ("from now on, always use short formats"), or the same behavior being repeated across several separate occasions. Promotion should also record a confidence level and a source, so the fact can be reviewed or corrected later. And users should be able to see, edit, and delete what the agent has remembered about them, which connects to the responsible-AI and privacy topics from Chapter 41.