	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > LangChain > LangChain Conversation Memory

LangChain Conversation Memory

Author: Venkata Sudhakar

By default an LLM has no memory - every call is independent. If a customer tells your support chatbot "my order number is 4821" in one message and then asks "when will it arrive?" in the next, the bot has no idea which order they mean. Conversation memory solves this by maintaining the message history and including it in every subsequent call to the LLM. The model can then refer back to anything said earlier in the session, making the conversation feel natural and coherent rather than like talking to someone with amnesia.

LangChain provides several memory classes. ConversationBufferMemory keeps every message in full - simple and reliable for short conversations. ConversationBufferWindowMemory keeps only the last K message pairs - useful when you want to limit token usage on long chats. ConversationSummaryMemory uses the LLM itself to compress old messages into a running summary, keeping costs low while preserving the key facts from earlier in the conversation. For a typical customer support chatbot, buffer memory with a window of 10 pairs covers most sessions comfortably.

The below example builds a shopping assistant that remembers the customer name, their product interest, and their budget across multiple turns - no repetition needed from the customer.

from langchain_openai import ChatOpenAI
from langchain.memory import ConversationBufferMemory
from langchain.chains import ConversationChain
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_core.runnables import RunnableWithMessageHistory
from langchain_community.chat_message_histories import ChatMessageHistory

llm = ChatOpenAI(model="gpt-4o-mini", api_key="your-api-key", temperature=0.4)

# Prompt template with a placeholder for the conversation history
prompt = ChatPromptTemplate.from_messages([
    ("system",
     "You are a helpful shopping assistant for FreshMart, an online grocery store. "
     "Be warm and concise. Remember customer preferences mentioned earlier in the chat."),
    MessagesPlaceholder(variable_name="history"),
    ("human", "{input}")
])

chain = prompt | llm

# In-memory session store - in production use Redis or a database
sessions = {}

def get_history(session_id: str) -> ChatMessageHistory:
    if session_id not in sessions:
        sessions[session_id] = ChatMessageHistory()
    return sessions[session_id]

# Wrap chain with message history - history is injected automatically
chatbot = RunnableWithMessageHistory(
    chain,
    get_history,
    input_messages_key="input",
    history_messages_key="history"
)

def chat(session_id: str, message: str) -> str:
    response = chatbot.invoke(
        {"input": message},
        config={"configurable": {"session_id": session_id}}
    )
    return response.content

# Simulate a customer conversation - each turn remembers the previous ones
sid = "customer-priya-001"
print("Customer:", "Hi, my name is Priya")
print("Bot:", chat(sid, "Hi, my name is Priya"))
print()
print("Customer:", "I am looking for organic products")
print("Bot:", chat(sid, "I am looking for organic products"))
print()
print("Customer:", "My budget is around Rs 500 per week")
print("Bot:", chat(sid, "My budget is around Rs 500 per week"))
print()
print("Customer:", "What would you recommend for me?")
print("Bot:", chat(sid, "What would you recommend for me?"))

It gives the following output,

Customer: Hi, my name is Priya
Bot: Hello Priya! Welcome to FreshMart. How can I help you today?

Customer: I am looking for organic products
Bot: Great choice, Priya! We have a wide range of certified organic fruits,
vegetables, dairy, and grains. Any particular category you have in mind?

Customer: My budget is around Rs 500 per week
Bot: Perfect - Rs 500 a week works well for organic basics. You could cover
seasonal vegetables, one or two fruits, and some pantry staples comfortably.

Customer: What would you recommend for me?
Bot: Based on your interest in organic products and a Rs 500 weekly budget,
here is a good starting basket for you, Priya:
- Organic seasonal vegetables (spinach, carrots, tomatoes) - Rs 180
- Organic bananas or apples - Rs 80
- Organic whole wheat flour or rice - Rs 120
- Organic dal or pulses - Rs 120
Total: Rs 500

# Notice: the bot remembered "Priya", "organic", and "Rs 500"
# without the customer repeating any of it in the final question

It gives the following output,

Messages stored in session: 8
Customer: Hi, my name is Priya
Bot: Hello Priya! Welcome to FreshMart. How can I help you today?
Customer: I am looking for organic products
Bot: Great choice, Priya! We have a wide range of certified organ...
Customer: My budget is around Rs 500 per week
Bot: Perfect - Rs 500 a week works well for organic basics. You c...
Customer: What would you recommend for me?
Bot: Based on your interest in organic products and a Rs 500 weekl...

# Full conversation stored - 4 turns = 8 messages
# Each new message is automatically appended by RunnableWithMessageHistory

Each session_id creates a completely isolated conversation - customer A and customer B never see each other messages. In production, generate a unique session ID when a customer starts a chat (UUID tied to their login session), store the ChatMessageHistory in Redis so it survives server restarts, and set a TTL so old sessions are cleaned up automatically. For very long conversations (more than 20 turns), switch to ConversationSummaryMemory to keep costs manageable while preserving key facts like the customer name, order number, and stated preferences.

Send your comments, suggestions or queries regarding this site to [email protected].