05. Add memory to Agent

Currently, chatbots cannot remember their past interactions themselves, so there are limits to conducting consistent multi-turn conversations.

To solve this in this tutorial memory Add.

Reference

This time pre-built ToolNode Wow tools_condition Use it.

1. ToolNode : Node for tool call

2. tools_condition : Condition branch based on whether the tool is called

Our chatbots can now use tools to answer user questions, but of previous interaction context I don't remember. This limits the ability to conduct multi-turn conversations.

LangGraph has persistent checkpointing Solve this problem.

When compiling graphs checkpointer When providing and calling the graph thread_id If you provide, LangGraph After each step Save status automatically To. same thread_id When you call the graph again using, the graph loads the saved state so that the chatbot can continue the conversation at the point where it was previously stopped.

checkpointing Is much more powerful than LangChain's memory function. (Maybe you can check this naturally by completing this tutorial.)

Copy

# Configuration file for managing API keys as environment variables
from dotenv import load_dotenv

# Load API key information
load_dotenv()

Copy

 True 

Copy

# Set up LangSmith tracking. https://smith.langchain.com
# !pip install -qU langchain-teddynote
from langchain_teddynote import logging

# Enter a project name.
logging.langsmith("CH17-LangGraph-Modules")

Copy

 Start tracking LangSmith. 
[Project name] 
CH17-LangGraph-Modules 

But before going too far, to enable multi-turn conversations checkpointing Let's add it.

MemorySaver Generates checkpointer.

Copy

from langgraph.checkpoint.memory import MemorySaver

# Create a memory store
memory = MemorySaver()

Reference

In this tutorial in-memory checkpointer Use.

But in the production phase, this SqliteSaver or PostgresSaver You can change to and connect to your own DB.

Copy

from typing import Annotated
from typing_extensions import TypedDict
from langchain_openai import ChatOpenAI
from langchain_teddynote.tools.tavily import TavilySearch
from langgraph.graph import StateGraph, START, END
from langgraph.graph.message import add_messages
from langgraph.prebuilt import ToolNode, tools_condition


########## 1. State Definition ##########
# State Definition
class State(TypedDict):
    # Add a message list comment
    messages: Annotated[list, add_messages]


########## 2. Tool definition and binding ##########
# Initialize tool
tool = TavilySearch(max_results=3)
tools = [tool]

# LLM Initialization
llm = ChatOpenAI(model="gpt-4o-mini")

# Combining tools and LLM
llm_with_tools = llm.bind_tools(tools)


########## 3. add note ##########
# Defining a chatbot function
def chatbot(state: State):
    # Calling and returning messages
    return {"messages": [llm_with_tools.invoke(state["messages"])]}


# Create a state graph
graph_builder = StateGraph(State)

# Add a chatbot node
graph_builder.add_node("chatbot", chatbot)

# Creating and adding tool nodes
tool_node = ToolNode(tools=[tool])

# Add tool node
graph_builder.add_node("tools", tool_node)

# Conditional Edge
graph_builder.add_conditional_edges(
    "chatbot",
    tools_condition,
)

########## 4. add edge ##########

# tools > chatbot
graph_builder.add_edge("tools", "chatbot")

# START > chatbot
graph_builder.add_edge(START, "chatbot")

# chatbot > END
graph_builder.add_edge("chatbot", END)

Finally, provided checkpointer Compile graphs using

Copy

# Compile the graph builder
graph = graph_builder.compile(checkpointer=memory)

Graph connectivity LangGraph-Agent Same as

Just, what was added this time was that the graph handled each node State Just checkpointing.

Copy

from langchain_teddynote.graphs import visualize_graph

# Graph visualization
visualize_graph(graph)

RunnableConfig Settings

RunnableConfig Define recursion_limit and thread_id Set.

• recursion_limit : Maximum number of nodes to visit. More than that, RecursionError

• thread_id : Thread ID setting

thread_id Is used to separate conversation sessions. In other words, the storage of memory thread_id It is done individually according to.

Copy

from langchain_core.runnables import RunnableConfig

config = RunnableConfig(
    recursion_limit=10,  # Visit up to 10 nodes, more than that will result in RecursionError
    configurable={"thread_id": "1"},  # 스레드 ID 설정
)

Copy

# First question
question = (
    "My name is `Teddy Note`. I run a YouTube channel. Nice to meet you."
)

for event in graph.stream({"messages": [("user", question)]}, config=config):
    for value in event.values():
        value["messages"][-1].pretty_print()

Copy

 ================================== Ai Message ================================== 

Hello, Teddy Note! Nice to meet you. I'm running a YouTube channel, so it's cool! What content are you mainly dealing with? 

Copy

# 이어지는 질문
question = "What did you say my name was?"

for event in graph.stream({"messages": [("user", question)]}, config=config):
    for value in event.values():
        value["messages"][-1].pretty_print()

Copy

 ================================== Ai Message ================================== 

Your name is Teddy Note. 

this time RunnableConfig of thread_id After changing, I'll ask if you remember the previous conversation.

Copy

from langchain_core.runnables import RunnableConfig

question = "What did you say my name was?"

config = RunnableConfig(
    recursion_limit=10,  # Visit up to 10 nodes, more than that will result in RecursionError
    configurable={"thread_id": "2"},  # Thread ID setting
)

for event in graph.stream({"messages": [("user", question)]}, config=config):
    for value in event.values():
        value["messages"][-1].pretty_print()

Copy

 ================================== Ai Message ================================== 

Sorry, there is no function to remember your name. If you haven't mentioned your name in the previous conversation, I don't know. I hope you give me your name! 

Snapshot: Check the saved State

So far, I have made several checkpoints in two different threads.

Checkpoint Is the current status value, its configuration, and to be processed next Node included.

Graph at a given setting state Anytime to check get_state(config) Call

Copy

from langchain_core.runnables import RunnableConfig

config = RunnableConfig(
    configurable={"thread_id": "1"},  # Thread ID setting
)
# Create a graph state snapshot
snapshot = graph.get_state(config)
snapshot.values["messages"]

Copy

 [HumanMessage (content=' My name is `Tedinot`. I am running a YouTube channel. Nice to meet you', additional_kwargs={}, response_metadata={}, id='0b4fb756-225d-49a0-a90b-acb39cb42d48'), AIMessage (content='Hello, Teddy Nice to meet you. I'm running a YouTube channel, so it's cool! '1':'Toks={'fusal': None}, response_metadata={'token_usage': {'completion_tokens': 45,'prompt_tokens': 105,'  lun-c9143f45-843c-478c-9360-47d2bce2207e-0', usage_metadata={'input_tokens': 105,'total_tokens': 150,' Nok:'Get',None={, respons_metadata}'token_usage': ={'completion_tokens': 12,'prompt_tokens': 165,'total_tokens':1  gpt-4o-mini-2024-07-18','system_fingerprint':'fp_0ba0d124f1','finish_reason':'stop','logprobs': None}, id='run-7cc53da-982d- response_metadata={}, id='c0c1ad0a-7c51-49ba-9aa8-fa125b501ab2'), AIMessage (content=' Your name is Tedinot.', additional_kwargs={'refus  None,'reasoning_tokens': 0},'prompt_tokens_details': {'audio_tokens': None,'cached_tokens':'gpt-4o-mini-2024-07  'stop','logprobs': None}, id='run-e65d423b-fb03-4237-8f51-15d8b56df860-0', usage_metadata={'input_tokens': 192,'  'stop','logprobs': None}, id='run-e65d423b-fb03-4237-8f51-15d8b56df860-0', usage_metadata={'input_tokens': 192,' 

snapshot.config You can check the config information set to output.

Copy

# Set config information
snapshot.config

Copy

 {'configurable': {'thread_id': '1','checkpoint_ns':'','checkpoint_id': '1ef99493-62b0-6020-8007-4fb8c65ebabc'}} 

snapshot.value You can check the state values stored so far by outputting.

Copy

# stored values
snapshot.values

Copy

 {'messages': [HumanMessage (content=' My name is `Tedinot`. I am running a YouTube channel. Nice to meet you', additional_kwargs={}, response_metadata={}, id='0b4fb756-225d-49a0-a90b-acb39cb42d48'), AIMessage (content='Hello, Teddy Nice to meet you. I'm running a YouTube channel, so it's cool! '1':'Toks={'fusal': None}, response_metadata={'token_usage': {'completion_tokens': 45,'prompt_tokens': 105,'  id='run-c9143f45-843c-478c-9360-47d2bce207e-0', usage_metadata={'input_tokens': 45,'total_tokens'Your name is Tedinot.', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 12,'prompt_tokens': 165,'tot  '4p','fp_0ba0d124f1','fp','fin','gout','gop','logprobs': None} response_metadata={}, id='c0c1ad0a-7c51-49ba-9aa8-fa125b501ab2'), AIMessage (content=' Your name is Tedinot.', additional_kwargs={'refus  {'audio_tokens': None,'reasoning_tokens': 0},'prompt_tokens_details': {'audio_tokens': None,'model_name':'gpt 0}})]}  'fp_0ba0d124f1','finish_reason':'stop','logprobs': None}, id='run-e65d423b-fb03-4237-8f51-15d8b56df860-0', usage_met  'fp_0ba0d124f1','finish_reason':'stop','logprobs': None}, id='run-e65d423b-fb03-4237-8f51-15d8b56df860-0', usage_met 

snapshot.next Output to find forward at this point Check the following nodes You can.

END Since it has been reached, the next node is output with an empty value.

Copy

# next node
snapshot.next

Copy

 () 

Copy

snapshot.metadata["writes"]["chatbot"]["messages"][0]

Copy

 AIMessage (content=' Your name is Tedinot.', additional_kwargs={'fusal': None}, response_metadata={'token_usage': 12,'promrun-e65d423b-fb03-4237-8f51-15d8b56df860-0', usage_metadata={'input_tokens': 192,'output_tokens': 12,'total_tokens': 204,'  reasoning': 0}}) 

To visualize metadata of complex structures display_message_tree Use functions.

Copy

from langchain_teddynote.messages import display_message_tree

# Metadata (output in tree format)
display_message_tree(snapshot.metadata)