Episodic, Semantic and Procedural Memory (in AI)

The classic human memory taxonomy

Human memory is not a single store but a set of distinct systems. Psychologists generally divide it first by duration: a short-lived working memory that holds and manipulates information for the current task, and long-term memory that persists for hours, years, or a lifetime. The dominant model of working memory comes from Alan Baddeley and Graham Hitch (1974), who replaced the older notion of a single short-term store with a system containing a central executive coordinating subsystems such as the phonological loop and the visuospatial sketchpad.

Long-term memory is itself split into types. Endel Tulving's influential 1972 distinction separated episodic memory (personally experienced events located in time and place) from semantic memory (general facts and concepts). Neuroscientist Larry Squire later organized memory into declarative (conscious, 'knowing that') and nondeclarative (unconscious, 'knowing how') branches, with episodic and semantic memory sitting under declarative, and procedural memory under nondeclarative. These three long-term types, episodic, semantic, and procedural, form the taxonomy most often imported into AI.

• Working memory: short-term, capacity-limited, holds the current context.
• Episodic memory: long-term, autobiographical events with time and place.
• Semantic memory: long-term, decontextualized facts and concepts.
• Procedural memory: long-term, learned skills and how-to knowledge.

Episodic memory: remembering events and context

Episodic memory is the record of specific events as they were experienced, complete with their context: when something happened, where, in what order, and how it felt. Tulving described it as supporting a kind of mental time travel, the ability to re-experience a past moment rather than merely know a fact about it. Recalling a particular conversation last Tuesday, or the sequence of steps taken to fix a problem yesterday, draws on episodic memory.

The defining feature is temporal and contextual specificity. Episodic memories are tied to a single occurrence and preserve narrative order. This makes them rich but also fragile and detailed, the substrate from which more general knowledge is later abstracted.

Semantic memory: facts and general knowledge

Semantic memory holds general knowledge about the world that is detached from any particular experience of learning it. Knowing that Paris is the capital of France, that water boils at 100 degrees Celsius at sea level, or that a friend prefers tea over coffee are semantic facts. The specific moment most of them were learned is typically lost, because the context has fallen away and only the fact remains.

Semantic and episodic memory are closely linked: repeated episodic experiences are gradually consolidated into stable semantic knowledge. A child who experiences many individual dogs eventually forms the general concept 'dog' without recalling each encounter. This consolidation, from specific events to durable facts, is a pattern AI systems try to reproduce.

Procedural memory: skills and how-to

Procedural memory is knowledge of how to perform actions: riding a bicycle, typing, or following a familiar workflow. It is classified as nondeclarative because it is expressed through performance rather than conscious recollection. Most people cannot verbalize the exact muscle movements involved in balancing on a bike, yet they execute them reliably. Procedural memory is acquired slowly through repetition and tends to be durable and long-lasting once learned.

The contrast with declarative memory is sharp. Declarative memory answers 'what'; procedural memory answers 'how'. The rules of a skill can be described (declarative) long before habits or the body actually master it (procedural), and the two can dissociate in cases of brain injury.

Mapping these types onto AI memory systems

AI researchers adopted this taxonomy to give language-model agents structured, persistent memory beyond a single context window. A frequently cited reference is the CoALA framework (Cognitive Architectures for Language Agents) by Sumers, Yao, Narasimhan, and Griffiths (2023), which gives a language agent a short-term working memory plus optional long-term memories: episodic for experience, semantic for knowledge, and procedural for code and skills. Frameworks for building agents commonly organize stored memory along these lines.

In practice each type is implemented differently. Episodic memory is usually a log of past interactions or task trajectories, often stored as embeddings in a vector database and retrieved by similarity. Semantic memory is a store of extracted, deduplicated facts about the user or domain, updated when new information supersedes the old. Procedural memory is frequently encoded outside any database entirely, in the model's weights, its system prompt, and its registry of tools or functions.

The Generative Agents work by Park and colleagues (2023) showed one influential pattern: a memory stream logs episodic observations, a retrieval function scores them by recency, importance, and relevance, and a reflection process periodically synthesizes those episodes into higher-level insights, effectively turning episodic records into reusable semantic knowledge.

• Episodic in AI: logged conversations and task histories, retrieved by relevance and recency.
• Semantic in AI: extracted facts and preferences, updated rather than duplicated.
• Procedural in AI: model weights, system prompts, and tool or function definitions.

Why most chatbots only simulate long-term memory

A base large language model has no persistent episodic or semantic store. What looks like memory inside a single chat is just the conversation transcript being re-fed into the context window on every turn. Once that window is exceeded or the session ends, the information is gone. The model's only durable knowledge lives in its frozen weights, which resemble procedural and semantic memory baked in at training time but cannot record anything specific to an individual user after training.

To create lasting memory, systems bolt an external store onto the model and retrieve relevant entries into the prompt as needed, an approach related to retrieval-augmented generation. This is genuine engineered memory, but it is not automatic: without an explicit episodic log, a fact-extraction step, and a retrieval policy, a chatbot only appears to remember while the context lasts. Distinguishing the three memory types clarifies exactly which capability a given product actually provides.

What a personal AI memory needs from each type

A system meant to act as a durable personal memory or 'second brain' needs all three types working together rather than any one alone. Episodic memory preserves the raw record of what was saved and discussed and when, so events can be recalled in context. Semantic memory distills stable facts and preferences from those events, so the system knows who a user is without replaying every interaction. Procedural memory governs how the system retrieves and presents information consistently over time.

Personal AI memory tools, sometimes called second-brain apps, typically let a user store material such as documents, photos, or notes and retrieve it later through natural-language queries, which is fundamentally an episodic-and-semantic retrieval problem layered on search. The taxonomy is a useful lens for evaluating any such tool: ask whether it actually logs events, abstracts durable facts, and applies consistent retrieval behavior, or whether it only mimics memory for the length of a conversation.