Unlike humans, LLMs have no persistent memory by default. Agent memory systems are what allow agents to maintain context and learn across sessions.

Types of Agent Memory

• In-context memory: Everything in the current context window — temporary, limited by token count
• External memory (vector store): Semantic search over past interactions and documents — persistent
• Episodic memory: Key facts and summaries stored and retrieved as needed
• Procedural memory: Learned workflows and instructions stored as retrievable documents

In Practice

Most production agents use a vector store (like Chroma or Pinecone) to store and retrieve relevant past context — the agent searches its memory before acting, just like a human would.

Reference:

LangGraph memory documentation

https://langchain-ai.github.io/langgraph/concepts/memory/