Agent Memory Types: Conversation, Episodic, Semantic

Why this matters

An agent's memory architecture determines what it remembers about your product across sessions and how it updates its beliefs — misunderstanding memory types leads to marketing strategies that only influence one-shot queries while missing persistent, session-spanning knowledge.

Try it yourself

1. In a live agent or chatbot with memory, mention your product name once early in a conversation, then ask an unrelated question 10 turns later — observe whether and how the agent surfaces the earlier mention.
2. Locate the memory configuration for one open-source agent framework (MemGPT, AutoGen, or similar) and identify which memory tier stores user-provided facts versus retrieved documents.
3. Write one 2-sentence product description optimized for semantic memory storage — prioritize entity-attribute-value triples over narrative prose and explain your choices.

Prompt your AI

Use these three in order. Each builds on the one before.

1. Basics & terminology

In one paragraph, explain the three main types of memory an AI agent can have — conversation, episodic, and semantic — and how each differs.

2. Why it works (the mechanism)

Walk me through how an agent with episodic memory would store and later retrieve a fact it learned about a product during a previous research session.

3. Advanced — application & what's next

Given an agent that uses both in-context conversation memory and a long-term semantic knowledge store, how would a factual error about your product propagate, and what would it take to correct it?

References

• https://arxiv.org/abs/2304.03442