Skip to main content

Conversation Buffer Memory

Overview​

The ConversationBufferMemory class in LangChain allows storing chat messages from a conversation and extracting them into a variable that can be injected into a prompt or model. This enables maintaining context over the course of a conversation.

There are two main ways to extract the chat history:

  • As a single string containing the concatenated messages
  • As a list of ChatMessage objects

The class has configurable parameters to control:

  • The memory key name returned
  • Whether messages are returned as strings or list
  • Which input and output keys are saved

Usage​

To use ConversationBufferMemory, first import and instantiate the class:

from langchain.memory import ConversationBufferMemory

memory = ConversationBufferMemory()

Then you can add chat messages from the user and AI:

memory.chat_memory.add_user_message("Hi there!")
memory.chat_memory.add_ai_message("Hello!")

To return the chat history as a single concatenated string:

memory.load_memory_variables({})

{'history': 'Human: Hi there!\nAI: Hello!'}

To return the chat history as a list of ChatMessage objects:

memory = ConversationBufferMemory(return_messages=True)

memory.load_memory_variables({})

{'history': [HumanMessage(content='Hi there!'),  
             AIMessage(content='Hello!')]}

Customizing Memory Keys​

You can customize the memory key returned, as well as the input and output keys saved:

# Change memory key
memory = ConversationBufferMemory(memory_key="chat_history")

# Change input key  
memory = ConversationBufferMemory(input_key="user_input")

# Change output key
memory = ConversationBufferMemory(output_key="bot_response")

Memory Size Management​

Since all messages are stored, the memory can grow very large. Some ways to manage the size:

  • Set a maximum number of messages with max_messages
  • Use a sliding window approach like ConversationBufferWindowMemory
  • Periodically clear old messages with chat_memory.clear()

Examples​

See the memory examples for more usage details.