Episodic vs Semantic Memory (AI)

Episodic and semantic memory are two categories of long-term memory that AI agent designers borrow from cognitive psychology, where Endel Tulving drew the distinction in 1972. Episodic memory holds specific events situated in time and tied to the agent's own experience — what happened, when, and how it turned out. Semantic memory holds decontextualized facts and concepts that are true regardless of when or how they were learned. Both sit in an agent's long-term store, outside the model's weights and outside the context window, and are retrieved into context as needed. The distinction sounds academic until you build an agent and discover that "remember things" is not one engineering problem but at least two.

Episodic memory in agents

For an AI agent, episodic memory records structured traces of past interactions: a timestamp, the task, the action taken, the conditions in play, and the outcome observed. Recalling "last time I tried this approach on this task, it failed for this reason" lets an agent learn from its own history and avoid repeating mistakes, a behaviour sometimes called experience-following. The engineering challenges are selection and decay: an agent cannot store every token of every session, so something must decide which episodes are worth keeping, how to summarize them without losing the load-bearing detail, and when an old episode has gone stale. Retrieval is typically similarity-based — embed the current situation, pull the most relevant past episodes — which means a bad embedding choice quietly degrades the agent's ability to recognize that it has been here before.

Semantic memory in agents

Semantic memory instead stores stable knowledge independent of any single episode: domain facts, product specifications, configuration details, user preferences, definitions. It answers "what is true" rather than "what happened to me." In practice it is often implemented as a curated knowledge base or vector store that the agent reads through retrieval, and its failure mode is different from episodic memory's: not forgetting, but going out of date. A fact ingested once ("this firmware supports that protocol") silently rots unless something revalidates it. In practice the two stores interact, just as Tulving later argued they do in humans — an agent consults general semantic knowledge while interpreting a specific episodic record, and repeated consistent episodes can be distilled into new semantic facts, which is how an agent's experience gradually becomes its knowledge.

A concrete example

Consider an agent helping run a small mining fleet. Semantic memory holds the stable facts: each machine's model, its hashboard count, the pool configuration, the rule that voltage is managed per hash domain. Episodic memory holds the history: on a given date, machine three dropped a chain after a power blip and a restart at reduced frequency recovered it. The next time similar symptoms appear, the agent's useful move — try what worked before — comes from the episodic store, while its ability to describe why comes from the semantic one. Design the stores separately and both retrievals get simpler and more reliable.

Why sovereignty applies

Distinguishing the two helps designers decide what to persist and how to retrieve it within an agent's broader memory system — and it clarifies what accumulates. An episodic store is a detailed diary of everything you asked, tried, and got wrong; a semantic store is a distilled model of your infrastructure, preferences, and work. Together they are among the most sensitive artifacts an AI system produces, which is why holding both on hardware you control is central to sovereign AI. D-Central documents these memory types for self-hosted deployments: the memory that makes an agent genuinely useful is exactly the memory you should not hand to someone else's cloud.