Agent memory
InMemorySaver · thread_id · Cross-call persistence

1Memoryless Agent: every invoke is a new conversation

By default,create_agentCreated Agentno memory. every timeagent.invoke()They are all completely independent: the Agent does not know what was said before, nor does it know the user's name.

agent = create_agent(deepseek_model())

# 第一次对话：告诉 Agent 我的名字和爱好
response1 = agent.invoke({
    "messages": [HumanMessage(content="Hello my name is Seán and my favourite colour is green")]
})
print(response1['messages'][-1].content)

# 第二次对话（独立的 invoke）：问 Agent 我的爱好
response2 = agent.invoke({
    "messages": [HumanMessage(content="What's my favourite colour?")]
})
print(response2['messages'][-1].content)

The second time it was called, the agent had no memory of the first conversation—it didn’t know the user’s name was Seán or that he loved the color green. every timeinvokeIt's all a new, context-free conversation.

2With memory vs without memory: intuitive comparison

3checkpointer: LangGraph’s state persistence mechanism

LangGraph’s memory mechanism is calledcheckpointer. How it works is simple:

• every timeagent.invoke()After execution, checkpointer will complete theState (including all message history)Save to storage
• Use the same one next timethread_idWhen calling, checkpointer firstLoad from storagePrevious State
• New user messages are appended to the existing message history, and the model can "see" all historical conversations.

4InMemorySaver: the simplest memory storage solution

existcreate_agentIncomingcheckpointer=InMemorySaver(), you can enable the memory function for Agent:

from langgraph.checkpoint.memory import InMemorySaver
from langchain.agents import create_agent

agent = create_agent(
    deepseek_model(),
    checkpointer=InMemorySaver(),   # 开启内存记忆
)

It’s not enough to have a checkpointer—it also needs to passconfigParameters passed inthread_id, telling the Agent which "thread" this conversation belongs to (that is, which user).

5thread_id: distinguish the conversation context of different users

thread_idIs the unique identifier of the conversation. same onethread_idmany timesinvokeShared message history; differentthread_idThe conversations are completely isolated and do not interfere with each other.

# config 字典：指定这次 invoke 属于哪个线程
config = {"configurable": {"thread_id": "1"}}

# 第一次调用：告知名字和爱好
response1 = agent.invoke(
    {"messages": [HumanMessage(content="Hello my name is Seán and my favourite colour is green")]},
    config   # 传入 config
)

# 第二次调用：同一个 config（同一个 thread_id）
response2 = agent.invoke(
    {"messages": [HumanMessage(content="What's my favourite colour?")]},
    config   # 同一个 config = 同一个对话历史
)

print(response2['messages'][-1].content)

6Complete multi-turn dialogue demonstration

Here is the complete two-turn conversation code, showing how memory works in an actual multi-turn conversation:

from langgraph.checkpoint.memory import InMemorySaver
from langchain.agents import create_agent
from pprint import pprint

# 创建有记忆的 Agent
agent = create_agent(
    deepseek_model(),
    checkpointer=InMemorySaver()
)

config = {"configurable": {"thread_id": "1"}}

# 轮次 1：建立上下文
response1 = agent.invoke(
    {"messages": [{"role": "user", "content": "Hello my name is Seán and my favourite colour is green"}]},
    config
)

# 轮次 2：利用上下文
response2 = agent.invoke(
    {"messages": [{"role": "user", "content": "What's my favourite colour?"}]},
    config
)

# 查看第二次 invoke 的完整消息历史
for msg in response2["messages"]:
    pprint(msg.content)
    print()

Note that the message history includesAll messages from two invokes(4 items in total). On the second call, the Agent saw the color green mentioned by Seán in the first conversation, so it answered correctly.

7Limitations of memory and production environment solutions

Limitations of InMemorySaver

• process memory: The data is stored in the memory of the Python process,Disappears after restarting the service
• Context window limits: As the conversation gets longer, the message history grows, eventually exceeding the context window of the LLM (8K-128K tokens for most models)
• Cannot be shared across processes: Multiple service instances cannot share the same memory

Comparison of production environment memory solutions