[LangGraph教程]LangGraph03——为聊天机器人添加记忆

上一篇[LangGraph教程]LangGraph02——使用工具增强聊天机器人中，我们为聊天机器人增加了调用外部工具的功能，但机器人现在还不能记得之前回答的上下文，这限制了其进行连贯的多轮对话的能力。

LangGraph 通过持久性检查点机制解决了这个问题。具体来说，如果在编译图时提供了一个 checkpointer，并在调用图时提供了一个 thread_id，LangGraph 会在每个步骤后自动保存当前状态。这样一来，当使用相同的 thread_id 再次调用图时，图会自动加载之前保存的状态，从而使聊天机器人能够从上次停止的地方继续对话。

这种检查点机制不仅支持多轮对话，还具备更强大的功能。例如，它可以用于错误恢复、人机协作工作流程、时间旅行交互等场景。

使用工具的机器人

代码如下，详细解释在[LangGraph教程]LangGraph02——使用工具增强聊天机器人

from typing import Annotated
from typing_extensions import TypedDict
from langgraph.graph.message import add_messages
from langgraph.graph import StateGraph# 1. 定义状态类
class State(TypedDict):messages:Annotated[list, add_messages]
# 2. 构建图    
graph_builder = StateGraph(State)# 3.定义工具节点 
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_openai import ChatOpenAIlocal_llm = ["deepseek-r1:8b","qwen2.5:latest"]llm = ChatOpenAI(model=local_llm[1], temperature=0.0, api_key="ollama", base_url="http://localhost:11434/v1")tool = TavilySearchResults(max_results=2)
tools = [tool]
llm_with_tools = llm.bind_tools(tools)def chatbot(state: State):return {"messages": [llm_with_tools.invoke(state["messages"])]}# 4. 把节点加入图 
# 第一个参数是节点名称，第二个是要调用的函数
from langgraph.prebuilt import ToolNodegraph_builder.add_node("chatbot", chatbot)tool_node = ToolNode(tools=[tool])
graph_builder.add_node("tools", tool_node)# 5. 把边加入图
from langgraph.prebuilt import tools_conditiongraph_builder.add_conditional_edges("chatbot",tools_condition,
)
graph_builder.add_edge("tools", "chatbot")
graph_builder.set_entry_point("chatbot")graph = graph_builder.compile()graph.invoke({"messages":"今天武汉天气怎么样"})from IPython.display import Image, display
display(Image(graph.get_graph().draw_mermaid_png()))

如何实现多轮对话

前面的步骤和之前的一样，这里不再赘述：

from typing import Annotated
from typing_extensions import TypedDict
from langgraph.graph.message import add_messages
from langgraph.graph import StateGraph# 1. 定义状态类
class State(TypedDict):messages:Annotated[list, add_messages]
# 2. 构建图    
graph_builder = StateGraph(State)# 3.定义工具节点 
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_openai import ChatOpenAIlocal_llm = ["deepseek-r1:8b","qwen2.5:latest"]llm = ChatOpenAI(model=local_llm[1], temperature=0.0, api_key="ollama", base_url="http://localhost:11434/v1")tool = TavilySearchResults(max_results=2)
tools = [tool]
llm_with_tools = llm.bind_tools(tools)def chatbot(state: State):return {"messages": [llm_with_tools.invoke(state["messages"])]}# 4. 把节点加入图 
# 第一个参数是节点名称，第二个是要调用的函数
from langgraph.prebuilt import ToolNodegraph_builder.add_node("chatbot", chatbot)tool_node = ToolNode(tools=[tool])
graph_builder.add_node("tools", tool_node)# 5. 把边加入图
from langgraph.prebuilt import tools_conditiongraph_builder.add_conditional_edges("chatbot",tools_condition,
)
graph_builder.add_edge("tools", "chatbot")
graph_builder.set_entry_point("chatbot")

在编译图时提供 checkpointer：checkpointer 是一个用于保存和加载状态的对象。它可以是一个简单的文件系统实现，也可以是一个更复杂的数据库实现。这里将检查点存在内存中，用MemorySaver实现

from langgraph.checkpoint.memory import MemorySavermemory = MemorySaver()graph = graph_builder.compile(checkpointer=memory)

在调用图时提供 thread_id：thread_id 是一个唯一标识符，用于区分不同的对话线程。通过使用相同的 thread_id，LangGraph 可以加载之前保存的状态，从而实现多轮对话。

这里配置了线程id为1

user_input = "Hi there! My name is Will."config = {"configurable": {"thread_id": "1"}}events = graph.stream({"messages": [{"role": "user", "content": user_input}]},config,stream_mode="values",
)
for event in events:event["messages"][-1].pretty_print()

输出如下：

================================Human Message=================================Hi there! My name is Will.
==================================Ai Message==================================Hello Will! How can I assist you today?

在每个步骤后，LangGraph 会自动调用 checkpointer 保存当前状态。当再次使用相同的 thread_id 调用图时，LangGraph 会自动加载之前保存的状态。

现在我们再次询问之前提到过的名称，看模型是否记得：

user_input = "Remember my name?"events = graph.stream({"messages": [{"role": "user", "content": user_input}]},config,stream_mode="values",
)
for event in events:event["messages"][-1].pretty_print()

回答如下：

================================Human Message=================================Remember my name?
==================================Ai Message==================================Of course! You mentioned your name is Will. How can I help you today, Will?

从回答中可以看到模型正确回答了这个问题。

现在，让我们换个线程id，看看模型是否记得：

# The only difference is we change the `thread_id` here to "2" instead of "1"
events = graph.stream({"messages": [{"role": "user", "content": user_input}]},{"configurable": {"thread_id": "2"}},stream_mode="values",
)
for event in events:event["messages"][-1].pretty_print()

从输出中可以看到，模型回答的并不对，说明不同线程之间不共享上下文

================================Human Message=================================Remember my name?
==================================Ai Message==================================Of course! You can always refer to me as Qwen. How can I assist you today?

最后，可以用get_state函数查看当前的上下文

snapshot = graph.get_state(config)
snapshot

结果如下:

StateSnapshot(values={"messages": [HumanMessage(content="Hi there! My name is Will.",additional_kwargs={},response_metadata={},id="bb8fc848-d778-4edb-86ed-af6c1b1b5cde",),HumanMessage(content="Hi there! My name is Will.",additional_kwargs={},response_metadata={},id="a38b9310-bab0-4838-b9ca-ef5656015012",),AIMessage(content="Hello Will! How can I assist you today?",additional_kwargs={"refusal": None},response_metadata={"token_usage": {"completion_tokens": 11,"prompt_tokens": 203,"total_tokens": 214,"completion_tokens_details": None,"prompt_tokens_details": None,},"model_name": "qwen2.5:latest","system_fingerprint": "fp_ollama","id": "chatcmpl-423","finish_reason": "stop","logprobs": None,},id="run-67c235d5-76ad-4287-a1bd-bb2a4729206f-0",usage_metadata={"input_tokens": 203,"output_tokens": 11,"total_tokens": 214,"input_token_details": {},"output_token_details": {},},),HumanMessage(content="Remember my name?",additional_kwargs={},response_metadata={},id="36e412ef-daa0-422e-b93c-25b9236985bc",),AIMessage(content="Of course! You mentioned your name is Will. How can I help you today, Will?",additional_kwargs={"refusal": None},response_metadata={"token_usage": {"completion_tokens": 20,"prompt_tokens": 227,"total_tokens": 247,"completion_tokens_details": None,"prompt_tokens_details": None,},"model_name": "qwen2.5:latest","system_fingerprint": "fp_ollama","id": "chatcmpl-217","finish_reason": "stop","logprobs": None,},id="run-649f5f0c-c2d2-4cb0-951d-8c62a56e717a-0",usage_metadata={"input_tokens": 227,"output_tokens": 20,"total_tokens": 247,"input_token_details": {},"output_token_details": {},},),]},next=(),config={"configurable": {"thread_id": "1","checkpoint_ns": "","checkpoint_id": "1f020170-d033-62b8-8006-dbfcb29ec140",}},metadata={"source": "loop","writes": {"chatbot": {"messages": [AIMessage(content="Of course! You mentioned your name is Will. How can I help you today, Will?",additional_kwargs={"refusal": None},response_metadata={"token_usage": {"completion_tokens": 20,"prompt_tokens": 227,"total_tokens": 247,"completion_tokens_details": None,"prompt_tokens_details": None,},"model_name": "qwen2.5:latest","system_fingerprint": "fp_ollama","id": "chatcmpl-217","finish_reason": "stop","logprobs": None,},id="run-649f5f0c-c2d2-4cb0-951d-8c62a56e717a-0",usage_metadata={"input_tokens": 227,"output_tokens": 20,"total_tokens": 247,"input_token_details": {},"output_token_details": {},},)]}},"step": 6,"parents": {},"thread_id": "1",},created_at="2025-04-23T07:46:22.313541+00:00",parent_config={"configurable": {"thread_id": "1","checkpoint_ns": "","checkpoint_id": "1f020170-b584-64d5-8005-f78ed589b2f8",}},tasks=(),
)

完整代码

from typing import Annotated
from typing_extensions import TypedDict
from langgraph.graph.message import add_messages
from langgraph.graph import StateGraph# 1. 定义状态类
class State(TypedDict):messages:Annotated[list, add_messages]
# 2. 构建图    
graph_builder = StateGraph(State)# 3.定义工具节点 
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_openai import ChatOpenAIlocal_llm = ["deepseek-r1:8b","qwen2.5:latest"]llm = ChatOpenAI(model=local_llm[1], temperature=0.0, api_key="ollama", base_url="http://localhost:11434/v1")tool = TavilySearchResults(max_results=2)
tools = [tool]
llm_with_tools = llm.bind_tools(tools)def chatbot(state: State):return {"messages": [llm_with_tools.invoke(state["messages"])]}# 4. 把节点加入图 
# 第一个参数是节点名称，第二个是要调用的函数
from langgraph.prebuilt import ToolNodegraph_builder.add_node("chatbot", chatbot)tool_node = ToolNode(tools=[tool])
graph_builder.add_node("tools", tool_node)# 5. 把边加入图
from langgraph.prebuilt import tools_conditiongraph_builder.add_conditional_edges("chatbot",tools_condition,
)
graph_builder.add_edge("tools", "chatbot")
graph_builder.set_entry_point("chatbot")from langgraph.checkpoint.memory import MemorySavermemory = MemorySaver()graph = graph_builder.compile(checkpointer=memory)user_input = "Hi there! My name is Will."config = {"configurable": {"thread_id": "1"}}events = graph.stream({"messages": [{"role": "user", "content": user_input}]},config,stream_mode="values",
)
for event in events:event["messages"][-1].pretty_print()user_input = "Remember my name?"events = graph.stream({"messages": [{"role": "user", "content": user_input}]},config,stream_mode="values",
)
for event in events:event["messages"][-1].pretty_print()# The only difference is we change the `thread_id` here to "2" instead of "1"
events = graph.stream({"messages": [{"role": "user", "content": user_input}]},{"configurable": {"thread_id": "2"}},stream_mode="values",
)
for event in events:event["messages"][-1].pretty_print()snapshot = graph.get_state(config)
print(snapshot)