Episodic, semantic, and procedural memory — keeping agents informed without blowing the context window.
Key vocabulary
Episodic memory — records of specific past events.
Semantic memory — general knowledge and facts about the domain.
Procedural memory — how to perform tasks: steps, tools, decisions.
Context window — the limited token space the agent can see at once.
Memory store — external storage (vector DB, database) persisting memory across sessions.
0 / 5 completed
1 / 5
A support agent remembers that a user reported a billing error last week and references it now. This is an example of:
Episodic memory = memories of specific events: who, what, when. Enables personalisation and continuity across sessions. Stored in a database or vector store, retrieved via similarity search.
2 / 5
An agent knows that “JWT tokens expire after a configurable TTL and must be refreshed using the /refresh endpoint.” This is:
Semantic memory = general knowledge, not tied to a specific event. Typically encoded in the system prompt or retrieved from a knowledge base.
3 / 5
When an agent’s context window approaches its limit, context compression involves:
Context compression: summarise earlier turns rather than truncating blindly. Ask the LLM to write a memory summary of older steps, then replace those steps with the summary.
4 / 5
A memory store in an agentic system refers to:
Memory store = external persistence. Common implementations: vector database (semantic retrieval), key-value store (structured facts), SQL database (episodic records).
5 / 5
An agent that knows how to file a support ticket — which fields, which API, in what order — is drawing on:
Procedural memory = knowing how to do something. Often encoded in the system prompt as instructions or retrieved from a task playbook memory store.