A durable AI agent should not treat every memory the same way. Some things are events — what happened, when, in what order. Others are facts — true regardless of when you learned them. Collapsing both into one bucket — a single vector store or a rolling chat log — is a common mistake in agent memory design. The cleaner model is older than any database.

Episodic memory stores specific events together with their temporal and spatial context; retrieval includes a sense of reliving the original experience. Semantic memory stores facts, concepts, and general knowledge stripped of personal context; retrieval feels like knowing rather than remembering.

These definitions come from Endel Tulving, who introduced the distinction in 1972. The split is still a useful design lens for engineers building persistent memory into AI agents.

TL;DR

• Tulving's 1972 framework separates time-stamped autobiographical events from context-free general knowledge.
• The two systems are interdependent: semantic structures guide encoding of new episodes, while repeated episodes gradually feed semantic memory.
• Current LLMs operate primarily as semantic stores; their context windows provide only transient, session-bound episodic capacity.
• Durable agent memory benefits from an explicit dual-store architecture plus a controlled consolidation process that decides when and how episodes become facts.

Tulving's Original Distinction (1972)

Endel Tulving (1927–2023) published the chapter "Episodic and Semantic Memory" in the volume Organization of Memory (Tulving & Donaldson, eds., Academic Press, pp. 381–403) [1]. He observed that existing memory research collapsed two different kinds of information under one heading.

Episodic memory records personal experiences tied to a specific time and place. A single exposure can create a lasting trace. When retrieved, the rememberer re-experiences elements of the original event — episodic memory answers "What happened?"

Semantic memory holds knowledge independent of acquisition context. "Lisbon is the capital of Portugal" exists without reference to when or where the fact was learned. Semantic retrieval is abstract and does not involve mental time travel — semantic memory answers "What is true?"

Tulving returned to the topic in 1985 with "Memory and consciousness" (Canadian Psychology, 26(1), 1–12). He linked episodic memory to autonoetic (self-knowing) consciousness — the ability to mentally project oneself backward or forward in time — and semantic memory to noetic (knowing) consciousness [2].

The framework was debated and refined for decades. Later work in Neuropsychologia (2020) and Memory & Cognition (2022) shows the field now treats the boundary as a gradient rather than a strict binary, yet the core engineering insight remains intact [3, 4].

Evidence of Distinct yet Interdependent Systems

Brain-damage cases support the separation. Damage to the hippocampus and nearby medial temporal lobe often causes anterograde amnesia. These patients cannot form new episodic memories, yet much of their earlier factual knowledge stays intact. The reverse also happens. Damage to the anterior temporal lobes can cause semantic dementia, which erodes factual knowledge while day-to-day event memory holds up.

Greenberg and Verfaellie (2010) examined these patterns and concluded that episodic and semantic memory are interdependent rather than isolated filing systems. Semantic knowledge scaffolds the interpretation and encoding of new experiences. At the same time, many semantic facts begin life as episodes that lose their contextual wrappers through repetition or abstraction. The influence runs in both directions, at both encoding and retrieval [5].

This interdependence matters for AI design. A purely semantic store cannot supply the contextual scaffolding that new episodes require. A purely episodic store cannot support rapid, decontextualized generalization. The right model is not two sealed boxes — it is two stores with different rules and a controlled path between them.

From Episodes to Facts: Semanticization and Consolidation

Episodic memories do not remain forever tied to their original context. Through repeated exposure, gist extraction, or integration into existing schemas, they gradually lose autobiographical detail and become semantic facts.

Cognitive scientists describe this process as semanticization. The biological analogy is systems consolidation: the hippocampus initially binds episodic details, then, over time and especially during sleep, transfers a more abstract version to neocortical networks. The Complementary Learning Systems (CLS) framework formalizes this division of labor [6].

For AI agents the timing of this transfer becomes an explicit engineering parameter. Consolidate too early and the system loses provenance and over-generalizes from limited evidence. Consolidate too late and retrieval remains slow and context-heavy. The correct cadence depends on task, reliability requirements, and the cost of maintaining raw event logs.

See the related article on schema formation for mechanisms that support gist extraction and integration.

Mapping Tulving's Split onto LLM Architectures

Today's large language models work mainly as semantic memory. Their weights hold huge amounts of context-free knowledge learned during pre-training. Pulling a fact from that store feels instant, and it carries no timestamp or original context.

The only episodic-like component is the context window. It records recent tokens in order, yet it lacks several properties that define human episodic memory:

• No durable temporal grounding across sessions
• No explicit source attribution
• No persistence beyond the current conversation
• No mechanism for selective consolidation

This thin episodic layer explains why agents built on plain LLM calls forget everything when the window clears, and cannot distinguish between a single observation and a repeated pattern.

Dual-Store Design for AI Agents

A practical architecture therefore maintains two separate stores with different invariants.

An event log (episodic store) appends records that preserve time, full context, source, and outcome. It never overwrites history. Retrieval can reconstruct sequences, calculate recency, and trace provenance.

A fact store (semantic store) keeps one clean, context-free version of each fact. It is built for fast similarity search, deduplication, and generalization. Its update rules differ from the event log's.

A consolidation step sits between them. It examines the event log for recurring patterns that exceed a chosen threshold and promotes a distilled version into the fact store. The threshold itself — minimum repetitions, confidence, schema match — becomes a tunable policy. A good fact store can still keep links back to the supporting episodes, so provenance survives even after context is stripped from the main retrieval path.

# Illustrative only — two stores with different rules.
class EpisodicStore:
    """Event log: preserves time, context, source."""
    def append(self, event, *, context, timestamp, source):
        ...  # never overwrites; ordered by time

class SemanticStore:
    """Fact store: context-free, deduplicated, fast to query."""
    def upsert(self, concept, fact):
        ...  # one canonical entry per concept

def consolidate(episodes, semantic_store, *, min_repeats):
    """Promote a pattern seen across enough episodes into a context-free fact."""
    for pattern in recurring_patterns(episodes, min_repeats=min_repeats):
        semantic_store.upsert(pattern.concept, pattern.gist)

This separation lets each component optimize for its purpose. The episodic store supplies grounded reasoning and auditability. The semantic store supplies speed and abstraction. The consolidation function decides when the cost of context is no longer justified.

Open questions remain: how to measure when an episode has been sufficiently "seen," how to preserve necessary source attribution after semanticization, and how to balance storage cost against retrieval latency. These are empirical and task-dependent.

Relation to Other Memory Mechanisms

Tulving's distinction does not replace other memory categories; it sits beside them. Working memory operates on the scale of seconds and feeds both episodic encoding and semantic updating. Schema formation supplies one of the main routes by which episodes become facts. Hopfield networks and modern associative memory models offer computational substrates for the rapid pattern completion that both stores require [7].

For a broader map of named memory systems see kinds of memory. For current approaches to measuring memory systems see evaluating agent memory. The 2026 landscape is surveyed in AI memory landscape 2026.

Common questions

What is the difference between episodic and semantic memory?

Episodic memory stores time-stamped personal experiences (what happened, when, where); semantic memory stores context-free facts and concepts. Tulving drew the line in 1972.

Do LLMs have episodic or semantic memory?

LLMs are almost entirely semantic — facts baked into parameters. Their only episodic layer is the context window, which has no durable temporal grounding or source attribution and is erased between sessions.

Should an AI agent keep episodic and semantic memory in separate stores?

Yes — they follow different rules. An event log preserves time, context, and source; a fact store strips context for fast generalization. Keeping them separate lets each do its job, with a consolidation step linking them.

What is memory consolidation (semanticization) in AI agents?

It is the deliberate step of turning repeated episodes into context-free facts — by repeated exposure, gist extraction, or schema integration — mirroring how the brain transfers memory from hippocampus to neocortex over time.

What did Endel Tulving contribute to memory research?

Tulving introduced the episodic-semantic distinction in 1972 and later added the autonoetic/noetic consciousness layer in 1985. He refined the framework across a career that spanned from 1927 to 2023.

Why are episodic and semantic memory interdependent?

Semantic knowledge scaffolds new episodes while many semantic facts originate in episodes. Damage to different brain regions produces dissociations, yet the systems influence each other at encoding and retrieval.

Related

• Kinds of memory — taxonomy overview that links into this deep dive on the episodic-semantic split.
• Working memory
• Schema formation
• Hopfield associative memory
• AI memory landscape 2026
• Evaluating agent memory

The same design questions Tulving framed for human cognition now confront every team building agents that must remember across sessions. Mnemoverse is a persistent-memory API for AI agents; the honest connection is narrow — it is one engineering answer to where the episodic/semantic line should sit, not a claim to reproduce human memory. By Edward Izgorodin. Free key: console.mnemoverse.com · Start here: Getting Started.

References

1. Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W. Donaldson (Eds.), Organization of Memory (pp. 381–403). Academic Press.
2. Tulving, E. (1985). Memory and consciousness. Canadian Psychology / Psychologie canadienne, 26(1), 1–12.
3. Renoult, L., & Rugg, M. D. (2020). An historical perspective on Endel Tulving's episodic-semantic distinction. Neuropsychologia, 139, 107366. https://doi.org/10.1016/j.neuropsychologia.2020.107366
4. De Brigard, F., Umanath, S., & Irish, M. (2022). Rethinking the distinction between episodic and semantic memory: Insights from the past, present, and future. Memory & Cognition, 50, 459–463. https://doi.org/10.3758/s13421-022-01299-x
5. Greenberg, D. L., & Verfaellie, M. (2010). Interdependence of episodic and semantic memory: Evidence from neuropsychology. Journal of the International Neuropsychological Society, 16(5), 748–753. https://pmc.ncbi.nlm.nih.gov/articles/PMC2952732/
6. McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological Review, 102(3), 419–457. https://doi.org/10.1037/0033-295X.102.3.419
7. Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79(8), 2554–2558.

---

Part of the Mnemoverse Library — research on AI memory for engineers.