Skip to main content

Memory in LangChain

Introduction​

LangChain provides a variety of memory types to store conversation context and enable your AI assistant to reference past interactions. This guide covers key concepts applicable across memory types and provides examples for working with different memory types.

Key Concepts​

Before using specific memory types, it's helpful to understand these core ideas:

Memory Keys​

Most memory types allow specifying a memory_key parameter that controls the key name returned by load_memory_variables(). For example:

memory = ConversationBufferMemory(memory_key="chat_history")
print(memory.load_memory_variables({}))  
# Returns: {'chat_history': 'Human: Hi there!'}

Input Keys vs Output Keys​

When saving context, most types take input and output dicts. By default, full dicts are saved. To control which keys are saved, use input_key and output_key parameters.

For example:

memory = ConversationBufferMemory(
   input_key="question",
   output_key="response"
)
# Only specified keys will be saved

Returning Messages vs String​

Some types return chat history as a string. To return a list of ChatMessage objects instead, set return_messages=True.

Memory Types​

ConversationBufferMemory​

Keeps a buffer of chat messages. Simplest memory type.

memory = ConversationBufferMemory()

# Add messages  
memory.add_user_message("Hi there!")
memory.add_ai_message("Hello!") 

print(memory.load_memory_variables({}))
# Returns string by default 
# {'history': 'Human: Hi there!\nAI: Hello!'}

# Return list of ChatMessages
memory = ConversationBufferMemory(return_messages=True)
print(memory.load_memory_variables({})) 
# {'history': [UserMessage(...), AIMessage(...)]}

EntityMemory​

Extracts and stores entities from chat history.

memory = EntityMemory()

# Entities will be extracted and stored
memory.add_user_message("My favorite food is pizza")
memory.add_ai_message("Pizza is delicious!") 

print(memory.entity_store.store) 
# {'favorite food': 'pizza'}

KeyValueMemory​

Stores key-value pairs explicitly set by user.

memory = KeyValueMemory()

# Explicitly store key-value 
memory.set("name", "John")   

print(memory.kv_store.store)
# {'name': 'John'}

Conclusion​

This covers some key memory types in LangChain. See the full list for more. Memory is essential for conversational AI to reference past interactions. LangChain provides tools to quickly incorporate different memory types into your assistants.