Episodic vs Semantic Memory in AI
Episodic memory stores specific events, what happened, when, and where. Semantic memory stores timeless facts and concepts, stripped of the moment they came from. In AI, episodic memory is a time-anchored log of past sessions; semantic memory is the distilled knowledge an agent reuses. The interesting part is the process that turns the first into the second.
Where the distinction comes from
The split is psychologist Endel Tulving's, introduced in 1972. Episodic memory holds "temporally dated episodes or events and their spatio-temporal relations." Semantic memory is a time-independent "mental thesaurus", facts, concepts, and rules, with no record of when you learned them. You remember your first day at a job (episodic); you know that Paris is the capital of France (semantic), without recalling the moment you learned it.
In neuroscience the two are physically distinct. The hippocampus encodes episodes; over time, systems consolidation transforms them, specific spatio-temporal detail is shed, and the gist migrates to neocortex as durable semantic knowledge. Episodic memories don't just sit there. They get promoted.
How it maps to AI memory
Modern AI memory systems borrow this taxonomy directly, because flat storage doesn't scale to a model of a person over months. A well-designed memory layer separates memory by type so it can recall the right kind of thing for a given question.
| Episodic | Semantic | |
|---|---|---|
| What it stores | A specific event, who, when, where, why | A timeless fact, preference, or rule |
| Time-anchored? | Yes | No |
| Example | "On June 2 the user chose Postgres over Mongo" | "The user prefers Postgres" |
| AI implementation | Event log / vector store, recency-weighted | Extracted facts, key-value or graph nodes |
| Neuroscience analog | Hippocampus | Neocortex |
| How they connect | → Consolidation / promotion abstracts recurring episodes into facts → |
Note the AI taxonomy usually adds two more types beyond these. Working memory is the in-context buffer the model reads during a single turn. Procedural memory is learned how-to, patterns and skills an agent reuses without re-deriving them. (We break down all four in types of AI memory.) Episodic and semantic are the two that mirror long-term human memory most directly, and the two consolidation moves between.
Why you need both, and the promotion between them
An agent with only episodic memory hoards transcripts: it can tell you what happened three sessions ago but can't generalize. An agent with only semantic memory knows facts but loses the narrative, it can't answer "what did we do last week" or trace why a decision was made. You need both, plus the link.
That link is consolidation: abstracting recurring episodes into stable semantic facts, then letting raw episodes decay by importance. It's the same move the brain makes when episodic detail fades into semantic gist. The canonical AI version is Stanford's Generative Agents (Park et al., 2023): memories are retrieved by a score combining recency, importance, and relevance, and when accumulated importance crosses a threshold the agent reflects, synthesizing raw observations into higher-level inferences. That reflection step is episodic-to-semantic promotion in miniature.
This is exactly why the promotion matters more than the storage. Two agents can both file an event correctly; only one notices that the same kind of event has happened five times and turns it into a rule.
How today's systems handle it
The major memory layers all touch this distinction, but they draw the line in different places:
| System | Episodic / semantic handling |
|---|---|
| Mem0 | Extracts facts from conversations and dedups them, when new info contradicts old, it updates the fact in place rather than duplicating. Strong at fact maintenance; promotion is agent-directed, not autonomous, and episodic recall is the weaker axis. |
| Zep / Graphiti | Splits memory into an explicit episode subgraph, a semantic entity subgraph, and a community subgraph, with bi-temporal tracking (event time and ingestion time). The closest competitor to a true episodic+semantic split, but graph-infrastructure-heavy, and promotion is entity extraction rather than salience-weighted decay. |
| MemGPT / Letta | Pages memory between in-context and external storage with an OS metaphor. The episodic/semantic separation exists, but consolidation is on-demand paging, not a learning process. |
A caution on benchmarks: vendor self-run scores in this space are contested and hard to compare, because judge, prompt, and model choices can swing accuracy by double digits. Treat any single headline figure skeptically. The one neutral data point worth anchoring to is from LongMemEval's own authors: commercial assistants show a meaningful accuracy drop across sustained, long-horizon interactions. That degradation is the entire reason this category exists.
Where iCog draws the line
Most systems store episodic or semantic well and bolt on the other. iCog's bet is the part in between. It treats episodic-to-semantic promotion as a first-class, scheduled algorithm, "dream consolidation", rather than an agent tool-call: recurring episodes are abstracted into semantic facts, patterns and skills are extracted, and memories decay by salience weight, mirroring hippocampus-to-neocortex consolidation instead of a one-shot LLM summarize.
The lineage is deliberate: ACT-R's activation-and-decay model plus Generative Agents' recency-importance-relevance retrieval and reflection. And the consolidation math is deterministic, the LLM only renders the final language, it doesn't decide the structure. Mem0 dedups facts; Zep builds a temporal graph; both store both memory types competently. iCog's distinct claim is making the learning between them the core algorithm rather than a side effect. (We unpack the mechanics in memory consolidation for AI agents.)
FAQ
What is the difference between episodic and semantic memory in AI? Episodic memory stores specific, time-anchored events ("on June 2 the user chose Postgres"). Semantic memory stores timeless facts and rules ("the user prefers Postgres"). Episodic answers what happened; semantic answers what's true.
Can an AI convert episodic memories into semantic memories? Yes, that's consolidation, or promotion. Recurring episodes are abstracted into stable facts, mirroring how human systems consolidation moves gist from hippocampus to neocortex. Generative Agents does a version of this when accumulated importance triggers a reflection step.
Why do AI agents need both? Episodic-only agents hoard transcripts and can't generalize; semantic-only agents know facts but lose the narrative and can't answer "what did we do last week." Both, plus the promotion between them, is what lets an agent actually learn.
Is RAG the same as semantic memory? No. RAG is a retrieval pattern, pulling relevant text into context at query time. Semantic memory is a type of memory that RAG can serve. You can retrieve episodic memories with the same machinery.
How is AI agent memory different from a vector database? A vector database is storage and similarity search. Agent memory adds the typing (episodic/semantic/procedural), the salience-weighted retrieval, and the consolidation that promotes and decays memories over time. The database is one component, not the system.
The distinction between events and facts is the easy part. The promotion between them, consolidation as a real, scheduled, salience-driven algorithm rather than a summarize call, is where a memory layer either learns or just stores. That's the line iCog is built on. Try iCog →