Tulving's Memory Model and What It Means for Digital Twins

AI can reason generally well, but there are no general people. Before AI can meaningfully help us, it needs to understand us individually: our history, our preferences, how those preferences change. That requires memory, and not the kind where you dump everything into a vector database.

In 1972, cognitive psychologist Endel Tulving proposed that human long-term memory isn't a single system. It's at least two: episodic memory (autobiographical experiences tied to time and place) and semantic memory (general facts and knowledge). This distinction turns out to be directly useful for building AI that maintains context about individual users.

This post covers Tulving's model, shows how it plays out in everyday cognition, and explains how it maps to AI memory architecture.

The foundations: episodic vs. semantic memory

Tulving argued that long-term memory comprises at least two distinct subsystems:

Episodic Memory: This is the autobiographical scrapbook of your life. It stores personal experiences tied to specific times, places, and contexts. When you recall an episodic memory, you engage in "autonoetic" consciousness—you mentally time-travel and relive the event. It's subjective, emotional, and contextual, like remembering the thrill of your first skydive, complete with the wind in your face and the date on the calendar.

Semantic Memory: In contrast, this is your internal encyclopedia of general knowledge. It holds facts, concepts, and meanings divorced from when or where you learned them. Recall here involves "noetic" consciousness—a sense of knowing without reliving. For instance, you know that Paris is the capital of France, but you might not remember the classroom where you first heard it.

The two systems interact. Episodic memories consolidate into semantic knowledge over time: repeated experiences turn into general rules. And semantic facts shape how we interpret new episodes. Neuroscience supports this: amnesia patients can lose personal memories while retaining world knowledge, or vice versa, suggesting the systems are genuinely separate.

For AI, this matters because a flat database can't capture the difference between "what happened" and "what I know." Digital twins need both to avoid treating every interaction identically.

How this plays out in everyday life

The two systems work together constantly. A few examples:

Planning a vacation

Say you're deciding on a trip to Italy. Your semantic memory kicks in with factual knowledge: Rome is home to the Colosseum; Italian cuisine features pasta and gelato; flights from New York average 8 hours. These are timeless facts you've accumulated from books, school, or Google.

But then episodic memory layers in the personal touch: You recall your 2018 visit to Florence, the crowded streets during summer, the taste of authentic tiramisu at a specific café, and the jet lag that ruined your first day. This reliving influences your choices—you opt for spring travel to avoid crowds, book a familiar airline, and prioritize food tours. Without episodic recall, your planning would be generic; without semantic facts, it'd lack foundation.

Over time, repeated trips consolidate episodic details into semantic rules. "Italy is best in spring or fall" starts as a personal anecdote and becomes generalized knowledge.

Professional networking

At a conference, you meet a colleague named Alex. Semantic memory provides the basics: Alex works at TechCorp, a company specializing in cloud computing; their stock rose 15% last quarter. This is objective knowledge from news or LinkedIn.

Episodic memory adds depth: You remember a heated debate at last year's event where Alex challenged your presentation on AI ethics, leading to a late-night discussion over drinks that forged a mutual respect. When reconnecting, you reference that "memorable debate in Chicago," evoking shared context and building rapport.

If Alex shares new facts (TechCorp's latest merger), your episodic system links them to the past event, updating your mental model. Episodic ties make factual information sticky and relevant.

Learning a skill

You're picking up guitar. Semantic memory stores the chords: A minor is fingered this way; scales follow mathematical patterns. You know Jimi Hendrix revolutionized electric guitar without recalling where you read it.

Episodic memory captures your practice sessions: The frustration of your first lesson on March 15th, the breakthrough jam with friends last weekend, or the blister from overplaying. These personal milestones motivate you and refine technique through "mental replay."

During sleep, consolidation abstracts episodic struggles into semantic expertise. You intuitively know the finger placements without reliving every fumble.

Applying Tulving's model to AI

A digital twin is an agent that maintains a user's knowledge, preferences, and history. Traditional AI memory treats all data the same way: vector databases and retrieval-augmented generation make no distinction between a fact about the world and a personal experience. Tulving's model provides the architecture for separating these.

Fidelius uses a four-network architecture mapped to Tulving's model:

Episodic Mapping: The Experience Network stores first-person actions and conversations with full temporal-spatial context, akin to human episodes. It captures "what happened when, where, and with whom," enabling the AI to "relive" user interactions.

Semantic Mapping: The World Network holds decontextualized facts about the external world, like "The Eiffel Tower is 330 meters tall" or "Python 3.12 released in October 2023." This is pure knowledge without personal ties.

Two additional networks, Opinion (for judgments) and Observation (for patterns), build on this base. During interactions, the system recalls from both episodic and semantic stores, reconstructs context, and updates through reflection.

AI Examples

Vacation planning in practice

User Query: "Help me plan a trip to Italy."

Semantic Recall: The AI pulls from World Network: Facts like Rome's landmarks, average costs, visa rules. Response: "Italy's high season is June-August; flights from your location (US) take 8-10 hours."

Episodic Recall: From Experience Network: Past interactions show you hated crowds in a 2023 Paris trip (stored as "User expressed frustration with tourist crowds in summer 2023 episode"). The AI "relives" this: "Remembering your Paris trip last year, where crowds ruined the Eiffel Tower visit, I'd suggest April for milder weather and fewer tourists."

Interplay in AI: Like human consolidation, Fidelius's sleep cycle clusters episodes and distills them into semantic insights (e.g., "User prefers off-peak travel"), updating the World Network for future use.

Professional advice in practice

User Query: "What do I need to know about networking with Alex from TechCorp?"

Semantic Recall: World Network: "TechCorp specializes in cloud solutions; recent merger with DataInc boosted stock by 15%."

Episodic Recall: Experience Network: "In episode from conference-2024, user debated AI ethics with Alex, leading to positive follow-up email." The AI reconstructs: "Based on our discussion after your ethics presentation last April, Alex appreciated your balanced view. Reference that to break the ice."

Interplay in AI: Observation Network synthesizes patterns (e.g., "User builds rapport through intellectual debates"), influencing Opinion Network judgments like "High confidence: User values evidence-based networking".

Skill development in practice

User Query: "Teach me guitar basics."

Semantic Recall: World Network: Chord diagrams, scale theory, historical notes on instruments.

Episodic Recall: Experience Network: "From session-2025-01-10, user struggled with barre chords but succeeded after 30 minutes practice." Response: "Last time, on January 10th, you nailed the F chord after focusing on finger strength. Let's build on that with A minor today."

Interplay in AI: Use ontology to classify facts (e.g., episodic as action-oriented), while sleep consolidation prunes weak memories and reinforces skills, mimicking human practice-to-expertise transition.

What this enables

Episodic memory makes responses specific to the user rather than generic. Semantic storage avoids redundancy. Source tracking reduces hallucinations by keeping provenance attached to facts. Consolidation handles long contexts without token overload.

As digital twins integrate with more tools and interfaces, the privacy questions around episodic data get harder. How much personal history should an AI store? Who controls it? These aren't hypothetical: they're design decisions that shape whether people trust these systems.

Tulving's episodic-semantic split is more than a psychology concept. It's a practical architecture for building AI that maintains context across interactions instead of treating every conversation as the first one.