Episodic vs Semantic vs Procedural AI Memory

Episodic Memory: What Happened

Episodic memory records specific events, interactions, and experiences with their temporal and situational context. In human terms, it is your memory of what happened at yesterday's meeting, what you ate for lunch last Tuesday, or what you discussed with a colleague in the hallway. Each episodic memory is anchored to a time, a place, and a set of participants.

In AI systems, episodic memory stores records of specific conversations and interactions. "The user reported a bug in the payment module on March 5th" is an episodic memory. "During the architecture review, the team decided to switch from MongoDB to PostgreSQL" is an episodic memory. Each record includes what happened, when it happened, and what the context was.

Episodic memory is essential for applications where history matters. Customer support systems need to recall previous ticket interactions. Project management assistants need to remember what was decided in previous meetings. Debugging assistants need to recall what was tried before and what the outcomes were. Without episodic memory, the AI has no timeline, no ability to reference past events, and no way to say "last time we tried X and it did not work."

The challenge with episodic memory is volume. Every conversation generates episodic memories, and they accumulate quickly. Raw storage of every interaction is unsustainable for long-term use. This is where consolidation becomes critical: the process of distilling lasting knowledge from episodes and archiving or deleting the raw event records.

Semantic Memory: What You Know

Semantic memory stores facts, concepts, and general knowledge independent of when or where they were acquired. In human terms, it is knowing that Paris is the capital of France, that water freezes at zero degrees, or that your project uses React. You know these facts without necessarily remembering the moment you learned them.

In AI systems, semantic memory stores user facts, project information, configuration details, and domain knowledge. "The user prefers Python over JavaScript" is a semantic memory. "The production database runs PostgreSQL 16" is a semantic memory. "The API rate limit is 100 requests per minute" is a semantic memory. These facts are stable, reusable, and context-independent.

Semantic memory is the most common and most valuable type for most AI applications. It is compact (one fact per memory rather than an entire event transcript), highly reusable (the same fact is relevant across many conversations), and easy to retrieve (fact-based queries have clear semantic similarity with stored facts).

The relationship between episodic and semantic memory is critical. Semantic memories are typically extracted from episodic ones. After a conversation where the user mentions they use PostgreSQL, the episodic record ("in the March 5th conversation, the user said they use PostgreSQL") gets consolidated into a semantic fact ("the project uses PostgreSQL"). The episodic record can then be archived while the semantic fact persists as active, retrievable knowledge.

Procedural Memory: How to Do It

Procedural memory stores learned skills, workflows, and behavioral patterns. In human terms, it is knowing how to ride a bicycle, how to type on a keyboard, or how to debug a race condition. You do not recall learning these skills explicitly; you just know how to do them.

In AI systems, procedural memory is the least developed of the three types, but it is emerging in agentic applications. Procedural memory stores learned sequences of actions: "when the user reports a slow query, first check the query plan, then look at index coverage, then check table statistics." It stores response patterns: "when explaining technical concepts to this user, use analogies from their domain (financial services)." It stores decision heuristics: "when choosing between two approaches, prioritize the one that maintains backward compatibility."

Procedural memory is harder to extract automatically than semantic or episodic memory because it requires identifying patterns across multiple interactions rather than extracting facts from individual conversations. An AI that helped the user debug five different performance issues might develop procedural knowledge about the user's debugging preferences, but extracting that pattern requires analysis across all five episodes.

Adaptive Recall's reflect tool supports procedural memory extraction. It synthesizes observations across recent interactions, identifying recurring patterns, learned preferences, and effective strategies. These reflections become procedural memories that guide future interactions: "this user prefers code examples over conceptual explanations" or "always check the CI pipeline status before suggesting deployment."

How They Work Together

The three memory types form a complementary system. Episodic memory captures raw experience. Semantic memory distills lasting knowledge from those experiences. Procedural memory identifies patterns across both episodes and facts that guide behavior.

A practical example: A coding assistant has a conversation where the user debugs a slow database query (episodic memory). From that conversation, the system extracts the fact that the database uses PostgreSQL 16 with specific indexing conventions (semantic memory). After several similar debugging sessions, the system identifies a pattern: this user always wants to see the query plan first (procedural memory). Future interactions benefit from all three: the assistant knows the database technology (semantic), can reference the previous debugging session if relevant (episodic), and starts by showing the query plan without being asked (procedural).

Most current memory frameworks focus primarily on semantic memory, with some episodic support. Procedural memory is the frontier, and the systems that implement it well will create significantly more natural and efficient interactions. Adaptive Recall supports all three through its specialized tools: store captures episodic and semantic information, reflect generates procedural insights, and the knowledge graph connects entities across all memory types.

Choosing What to Store

Not every application needs all three types. The right mix depends on your use case.

A customer support bot primarily needs semantic memory (customer facts and preferences) and episodic memory (previous ticket history). Procedural memory helps but is not critical. A coding assistant benefits enormously from all three: semantic (project facts), episodic (debugging history), and procedural (learned workflows). A simple Q&A bot might only need semantic memory (domain knowledge) with no need for episodic or procedural types.

Start with semantic memory because it provides the highest return on the simplest implementation. Add episodic memory when conversation history and event timelines are important. Add procedural memory when you want the system to adapt its behavior based on learned patterns. This progression matches the order of increasing implementation complexity and the order of diminishing marginal returns for most applications.