Adding Memory to Chains

Introduction

Chains in LangChain can be made stateful by initializing them with a Memory object. This allows data to persist across multiple calls to the chain. The Memory module in LangChain provides various utilities for adding memory to chains.

Get Started

To get started with memory, you can initialize a chain like ConversationChain with a Memory object:

from langchain.chains import ConversationChain  
from langchain.memory import ConversationBufferMemory
  
conversation = ConversationChain(
  llm=chat,
  memory=ConversationBufferMemory()  
)

This will store the conversation history in the Memory object and make it available across runs of the chain.

ConversationBufferMemory

One simple memory type is ConversationBufferMemory. This keeps a buffer of the full conversation history. For example:

memory = ConversationBufferMemory()
memory.chat_memory.add_user_message("hi!") 
memory.chat_memory.add_ai_message("hello!")

The full buffer is then available to inject into prompts.

ConversationSummaryMemory

For longer conversations, ConversationSummaryMemory can be used to store a summary instead of the full history. This summarizes the conversation on the fly using an LLM.

For example:

memory = ConversationSummaryMemory(llm=GPT3)
# Messages get summarized
memory.save_context(...) 

# Summary returned 
memory.load_memory_variables({})

You can also initialize the memory directly from existing messages:

history = ChatMessageHistory()
# Add messages
memory = ConversationSummaryMemory.from_messages(llm, history) 

Conclusion

The LangChain Memory modules make it easy to add stateful memory to chains. Developers can leverage different storage options like buffering or summarization depending on their needs.