Contrast episodic and semantic long term memory with examples from a personal assistant agent

Flashcard·Medium·4.0 · 0·~30s·Asked atDroom Linkedin Mistral AI

Attempt it

Episodic memory holds specific past events with time and context: 'On 2026-04-12 the user asked about migrating from Pinecone to Turbopuffer and decided against it because of latency.' Semantic memory holds distilled, stable facts: 'The user prefers Python over TypeScript' or 'The user works at a 20-person ML startup.' Episodic memories are retrieved when the model needs to reference what happened; semantic memories are loaded as persona context every session. Mem0 and Zep both make this distinction explicit; Letta and LangGraph leave it to the application schema.

Your answer

Long-term memory in an LLM agent is not one thing. It splits cleanly into two kinds. **Episodic memory** holds specific past events tied to a time and context. Examples for a personal assistant: 'On 2026-04-12, the user asked about migrating from Pinecone to Turbopuffer and decided against it because of latency concerns.' 'On 2026-05-30, the user mentioned a sprint deadline of June 14.' These memories answer 'what happened?' They have timestamps, they reference specific decisions or events, and they are dense with context. **Semantic memory** holds distilled, stable facts about the user or domain. Examples: 'The user prefers Python over TypeScript.' 'The user works at a 20-person ML startup.' 'The user's primary timezone is IST.' These memories are time-invariant (or at least slow-changing), they describe attributes rather than events, and they are short and reusable. The two have different retrieval patterns. **Semantic memories** are loaded as persona context at the start of every session, they shape how the agent talks to this user. **Episodic memories** are retrieved on demand when the model needs to reference a specific past event ('what did we decide about vector stores last month?'). Mem0 and Zep both expose this split as first-class API constructs. Letta and LangGraph leave it to the application schema, but most teams arrive at the same distinction by necessity.

**Episodic stores events with time; semantic stores facts.** The distinction is borrowed from cognitive science but earns its keep in LLM systems because the two kinds have different shapes and different lifecycles. **Episodic memory entries** look like: `{event: 'User decided against Turbopuffer migration', date: '2026-04-12', reasoning: 'latency concerns from p99 benchmark', related_topics: ['vector-stores', 'pinecone', 'turbopuffer']}`. They are dense, time-stamped, and retrieved by similarity to a current question. Their value comes from preserving the *story* of what happened. When the user later asks 'remind me why we picked Pinecone', the episodic memory contains the actual reasoning, not just the conclusion. **Semantic memory entries** look like: `{fact: 'Prefers Python over TypeScript', confidence: 0.9, last_confirmed: '2026-05-15'}` or `{fact: 'Works at 20-person ML startup', confidence: 0.95}`. They are short, stable, and loaded into the system prompt as persona context. Their value is shaping the model's default behavior every session: pick Python when offering code, assume small-team constraints when discussing infrastructure. **Retrieval differs.** Semantic memory is eager, loaded at session start as a small persona block. Episodic memory is lazy, retrieved on demand via vector search over event embeddings when the current turn references the past. **Write policy differs.** Episodic writes are easier; you record what happened with timestamps and move on. Semantic writes are harder because the system must infer a stable trait from observations, often using an LLM to compress 'the user mentioned Python in 6 of the last 10 conversations' into 'prefers Python', and then maintaining a confidence score that can be revised. The two work together. A conversation often produces both: 'On 2026-06-01, the user shipped the auth refactor in Go' is episodic; the semantic update is bumping confidence on 'works on backend systems' or adding 'has Go

TL;DR

Episodic memory stores time-stamped events; semantic memory stores distilled stable facts. Episodic is retrieved on demand; semantic is loaded as persona context every session.

Memory aid

Easy to grasp

Think of what you remember about a friend. You remember specific moments, that birthday dinner two years ago when you both got food poisoning, and these come with dates and details. That is episodic. You also remember stable facts about them, they hate cilantro, they live in Berlin, they work as a vet. These are not tied to a particular event, just things that are true. That is semantic. An assistant needs both. The specific story of last month's Pinecone decision is episodic; 'the user always wants Python code, not TypeScript' is semantic. The two are stored, indexed, and used differently.

Key concepts

Context Engineering

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Cognitive science has carried the episodic versus semantic memory distinction since Tulving's work in the 1970s. LLM memory systems converged on the same split independently because the operational pressures are identical: events and facts have different shapes, lifecycles, and access patterns. Treating them the same produces a memory layer that is either too noisy (semantic entries that should be events) or too lossy (episodes flattened into bland generalizations).

A personal-assistant agent makes the distinction concrete. Some things the agent should know about the user are tied to specific moments, what we decided, when, why. Other things are stable traits, preferences, role, environment. The two combine to make the agent feel personalized without being noisy.

Definitions and example payloads

Episodic memory captures specific past events with temporal and contextual detail. The canonical shape is {event_id, timestamp, user_id, summary, full_context, related_topics, importance_score}. An example entry: {event_id: ep_8472, timestamp: '2026-04-12T14:32Z', summary: 'User decided against migrating from Pinecone to Turbopuffer', full_context: 'Discussed p99 latency benchmark showing 80ms vs 35ms at top-50 retrieval. User said cost savings would not offset the latency for their realtime use case.', related_topics: ['vector-stores', 'pinecone', 'turbopuffer'], importance_score: 0.7}. These memories tell the story.

Semantic memory captures distilled, time-invariant or slow-changing facts. The canonical shape is {fact_id, user_id, attribute, value, confidence, last_confirmed, source_episodes}. An example entry: {fact_id: sm_032, user_id: u_001, attribute: 'preferred_language', value: 'Python', confidence: 0.92, last_confirmed: '2026-05-15', source_episodes: [ep_4101, ep_5021, ep_5847]}. These memories tell what is true.

The source_episodes field is the audit trail: semantic facts should be traceable back to the episodic events that justified them. This becomes the foundation for trust ('the assistant says it knows X about me; how does it know?') and for revision ('this fact is no longer true; which episodes should I now down-weight?').

Retrieval patterns

Write policy: extract semantic from episodic, asynchronously

Decay, conflicts, and the temporal dimension

How frameworks land

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Real products, models, and research that use this idea.

Mem0 exposes episodic and semantic memory as first-class API types with separate extraction and retrieval paths.
Zep's temporal knowledge graph models episodic events as time-stamped graph edges and semantic facts as entity attributes.

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you implement the asynchronous extraction step that distills semantic facts from episodic memory?

Batch recent episodes, run an LLM with a structured-output schema asking 'what stable preferences are supported by repeated evidence here?', merge into existing semantic store with confidence updates.

1 more follow-up an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Storing every interaction as either pure episodic or pure semantic, missing that the same conversation often generates both kinds and they have different lifecycles.

60 second bullets to scan on the way to the call.

Define episodic memory with a concrete event-shaped example
Define semantic memory with a concrete attribute-shaped example

Primary sources. Browse if you want the original framing.