Logo of AI Product Engineer
AI Product Engineer
Tutorial Image: LangGraph Tutorial: Unit 1: Graph Basics & State Management - Exercise 1.2 - "Message Memory"

LangGraph Tutorial: Unit 1: Graph Basics & State Management - Exercise 1.2 - "Message Memory"

Extend Exercise 1.1 to create a more sophisticated conversation manager that: 1. Maintains a sliding window of recent messages 2. Generates summaries of longer conversations 3. Includes metadata with messages 4. Follows the same conversation flow (Hello -> How are you? -> Goodbye)

Rod Rivera

🇬🇧 Chapter

#tutorial

🎯 What You'll Learn Today

Unit 1: Graph Basics & State Management

Exercise 1.2 - "Message Memory"

Try It Yourself

Open In ColabDownload Notebook

📢 Joint Initiative

This tutorial is part of a collaboration between AI Product Engineer and Nebius Academy.

Objective

Extend Exercise 1.1 to create a more sophisticated conversation manager that:

  1. Maintains a sliding window of recent messages
  2. Generates summaries of longer conversations
  3. Includes metadata with messages
  4. Follows the same conversation flow (Hello -> How are you? -> Goodbye)

Requirements

  1. State Management:

    • Extend the State type to include summary and window_size
    • Track message metadata (timestamps, roles)
    • Maintain conversation history with sliding window

  2. Node Implementation:

    • llm_response: Generate appropriate responses with metadata
    • message_windowing: Implement sliding window logic
    • summary_generation: Create summaries for longer conversations
    • should_end: Determine when to end the conversation

  3. Graph Structure:

    • Create a pipeline of nodes for message processing
    • Implement proper conversation ending condition
    • Connect nodes in the correct order
from typing import Annotated, TypedDict

from langchain_core.messages import BaseMessage
from langgraph.graph import StateGraph
from langgraph.graph.message import add_messages
class State(TypedDict):
    Extended state for message memory management.
    Attributes:
        messages: List of conversation messages with metadata
        summary: Current conversation summary
        window_size: Maximum number of messages to keep
    messages: Annotated[list[BaseMessage], add_messages]
    summary: str
    window_size: int
def llm_response(state: State) -> State:
    Generate responses while maintaining conversation context.
    Args:
        state: Current conversation state with message history
    Returns:
        Updated state with new message and preserved context
    Notes:
        - Handle empty initial state
        - Add appropriate metadata to messages
        - Follow the Hello -> How are you? -> Goodbye flow
    # TODO: Implement response generation with metadata
    pass
def message_windowing(state: State) -> State:
    Maintain a sliding window of recent messages.
    Args:
        state: Current conversation state
    Returns:
        State with message history trimmed to window size
    # TODO: Implement sliding window logic
    pass
def summary_generation(state: State) -> State:
    Generate summaries for longer conversations.
    Args:
        state: Current conversation state
    Returns:
        State with updated summary if needed
    # TODO: Implement summary generation
    pass
def should_end(state: State) -> bool:
    Determine if the conversation should end.
    Args:
        state: Current conversation state
    Returns:
        True if the conversation should end, False otherwise
    # TODO: Implement ending condition
    pass
# Initialize the graph
graph_builder = StateGraph(State)

TODO: Add your nodes Hint: You need llm_response, message_windowing, and summary_generation

TODO: Add your edges Hint: Consider the order of operations:

  1. Generate response
  2. Apply windowing
  3. Update summary
  4. Decide whether to continue or end

TODO: Compile the graph graph = None # Replace with proper compilation

# Default input for testing
default_input = {"messages": [], "summary": "", "window_size": 3}
# Make variables available for testing
__all__ = ["default_input", "graph"]