5 LangChain Entity Extraction and Memory (EntityStore)

Most LangChain memory implementations store what was said. Entity memory stores what was learned — specifically, it extracts and maintains a structured understanding of the people, organizations, places, and concepts that appear in a conversation.

The practical difference matters when a user says "remind me what you know about John" ten turns into a conversation. With buffer memory, you hope the context window still includes the early messages where John was introduced. With entity memory, you have a dedicated store that tracks everything said about John from the first mention.

This guide covers five patterns for entity extraction and memory in LangChain: the basic setup, custom entity stores, Redis-backed persistence, LCEL chain integration, and a production customer service implementation.

For related memory patterns, see AI agent memory and planning and Build AI chatbot Python.

How Entity Memory Works Under the Hood

When you use ConversationEntityMemory, each conversation turn goes through two LLM calls:

1. Entity extraction: An LLM reads the recent conversation history and identifies entities (people, orgs, places, things) mentioned in the new message.
2. Entity summarization: For each identified entity, the LLM updates its stored description by synthesizing what was known before with what was just learned.

The entity store holds key-value pairs like:

"Alice" -> "Alice is a software engineer at Acme Corp. She prefers Python and is working on a RAG system."
"Acme Corp" -> "Technology company where Alice works. Focus on enterprise AI solutions."

These summaries are injected into the next turn's context window, giving the LLM grounded information about known entities without requiring the full conversation history.

Pattern 1: Basic ConversationEntityMemory

from langchain.memory import ConversationEntityMemory
from langchain_openai import ChatOpenAI
from langchain.chains import ConversationChain
from langchain_core.prompts import PromptTemplate

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0.7)

# Initialize entity memory with in-memory store
memory = ConversationEntityMemory(
    llm=llm,
    return_messages=True
)

# Custom prompt that uses entity context
prompt = PromptTemplate(
    input_variables=["entities", "history", "input"],
    template="""You are a helpful assistant with an excellent memory.

Known information about entities in this conversation:
{entities}

Conversation history:
{history}