Agent with Memory: Equip Agent with persistent memory
checkpointer · MemorySaver · Thread ID

1Problem: Stateless Agent cannot conduct multiple rounds of dialogue

In the previous lecture (1-4), we built a complete ReAct Agent: it can call tools in a loop until the problem is solved. But it has a fatal flaw——Each invoke() is a completely new conversation, don’t remember anything said before.

Why is this happening? State life cycle

State in LangGraph isAn object temporarily created each time invoke(). The execution process is as follows:

Essentially, this is because LangGraph's graph implementation ispure functional——Input comes in, output goes out, no side effects are retained. This is a good thing in most cases (predictable, testable), but multi-turn dialogue scenarios require state persistence.

2Solution: checkpointer checkpoint mechanism

Checkpointer(checkpointer) is the persistence layer of LangGraph. Its responsibility is: after each node of the graph is executed,Automatically serialize and save the current State to the storage backend。

What does checkpointer do (three things)

MemorySaver: The simplest checkpointer implementation

from langgraph.checkpoint.memory import MemorySaver

# MemorySaver 把所有 checkpoint 存在 Python 进程内存里
# 本质上是一个字典：{thread_id: {step: state_snapshot}}
memory = MemorySaver()

# 编译时把 checkpointer 传进去
react_graph_memory = builder.compile(checkpointer=memory)

MemorySaverUses an in-memory dictionary to store all checkpoints, zero configuration, out-of-the-box, suitable for development phases and unit testing. The disadvantage is that all data is lost after the process is restarted.

3Thread ID: distinguish between different conversation sessions

An Agent may serve multiple users at the same time, and the conversation history of each user must be isolated.Thread IDIt is the "conversation session ID", which is equivalent to the primary key in the database.

Complete conversation history of three threadsIndependent storage, no interference with each other. Alice's "that" is 7, Bob's "that" is 25, and Carol's "that" is 42.

import uuid

# 方式一：固定字符串（测试/demo 时常用）
config = {"configurable": {"thread_id": "1"}}

# 方式二：UUID（生产环境，每个用户 session 一个）
thread_id = str(uuid.uuid4())
config = {"configurable": {"thread_id": thread_id}}

# 方式三：与业务 ID 绑定（最常见的实践）
config = {"configurable": {"thread_id": f"user_{user_id}_session_{session_id}"}}

4Core code: only one parameter missing

Adding memory to Agent, the code changes are extremely small——Just pass in checkpointer when compile(), all the rest of the code remains unchanged.

from langgraph.graph import (
    StateGraph, MessagesState,
    START, END
)
from langgraph.prebuilt import (
    ToolNode, tools_condition
)

# 无 checkpointer
react_graph = builder.compile()

# 每次 invoke 都是全新对话
messages = react_graph.invoke(
    {"messages": messages}
    # 没有 config，无法区分对话
)

from langgraph.graph import (
    StateGraph, MessagesState,
    START, END
)
from langgraph.prebuilt import (
    ToolNode, tools_condition
)
from langgraph.checkpoint.memory import MemorySaver

memory = MemorySaver()

# 传入 checkpointer ← 唯一改动！
react_graph_memory = builder.compile(
    checkpointer=memory
)

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

# 必须传 config，带上 thread_id
messages = react_graph_memory.invoke(
    {"messages": messages},
    config   # ← 告诉图用哪个 thread
)

Full code: two-turn conversation example

from langchain_core.messages import HumanMessage
from langgraph.checkpoint.memory import MemorySaver

# ─── 初始化 Checkpointer ───
memory = MemorySaver()
react_graph_memory = builder.compile(checkpointer=memory)

# ─── 配置：使用 thread_id = "1" ───
config = {"configurable": {"thread_id": "1"}}

# ─── 第一轮对话 ───
messages = [HumanMessage(content="Add 3 and 4.")]
result = react_graph_memory.invoke({"messages": messages}, config)
# LLM 调用 add(3, 4) → ToolMessage("7") → AIMessage("The sum is 7.")
# ✅ Checkpointer 把这整个 State 保存到 thread "1" 的 checkpoint

# ─── 第二轮对话：注意只传了新的问题 ───
messages = [HumanMessage(content="Multiply that by 2.")]
result = react_graph_memory.invoke({"messages": messages}, config)
# ✅ LangGraph 先加载 thread "1" 的历史，再追加新消息
# LLM 看到完整历史，正确理解 "that" = 7
# → 调用 multiply(7, 2) → ToolMessage("14") → AIMessage("The result is 14.")

print(result["messages"][-1].content)
# 输出："The result is 14."

5How to "connect" two invokes - complete life cycle

To truly understand the memory mechanism, you need to know what steps each invoke() goes through internally.

6Comparison demonstration: with memory vs without memory

The following visually shows what LLM actually received during the second round of invoke().messagesWhat a difference a list makes.

Complete message list after the second round of execution (with memory version)

After the two rounds of dialogue are executed, there are a total of messages in the State of thread "1"7 items：

# thread_id="1" 的最终 messages 列表：
[
    HumanMessage(content="Add 3 and 4."),
    AIMessage(content="", tool_calls=[{"name": "add", "args": {"a": 3, "b": 4}}]),
    ToolMessage(content="7", name="add"),
    AIMessage(content="The sum of 3 and 4 is 7."),
    HumanMessage(content="Multiply that by 2."),
    AIMessage(content="", tool_calls=[{"name": "multiply", "args": {"a": 7, "b": 2}}]),
    ToolMessage(content="14", name="multiply"),
    AIMessage(content="The result of multiplying 7 by 2 is 14."),
]

7checkpointer workflow diagram

The figure below shows the graph with checkpointer inSingle invoke() internalHow to automatically save State after each node.

8MemorySaver vs Production Grade Storage

MemorySaverOnly suitable for development stage. Production environments require a checkpointer implementation that can be persisted to disk or an external database.

All checkpointer implementations follow the same interface (BaseCheckpointSaver), so when switching from MemorySaver to PostgresSaver,Zero changes to business code, just replace the checkpointer object passed in at compile time.

# 开发阶段
from langgraph.checkpoint.memory import MemorySaver
graph = builder.compile(checkpointer=MemorySaver())

# 生产阶段（替换一行，其余代码不变）
from langgraph.checkpoint.postgres import PostgresSaver
with PostgresSaver.from_conn_string("postgresql://user:pass@host/db") as checkpointer:
    graph = builder.compile(checkpointer=checkpointer)
    # 同样的 invoke()，同样的 config，完全一致的行为

9The necessity of Config parameters

Even if you pass checkpointer, if it is not provided when invoke()config(or no thread_id), LangGraph willReport an error. This is a mandatory constraint by design.

# ❌ 错误：有 checkpointer 但没有 config
react_graph_memory.invoke({"messages": messages})
# 抛出 ValueError: "thread_id" is required in config["configurable"]
# when a checkpointer is used

# ✅ 正确：提供 config 和 thread_id
config = {"configurable": {"thread_id": "1"}}
react_graph_memory.invoke({"messages": messages}, config)

Why is it mandatory? Because there is no thread_id, checkpointer doesn't know at all:

• Where to load history from(The data of multiple users is mixed together, which one should be used?)
• Where should the new checkpoint be saved?(Save to which conversation session?)

Query history State

With checkpointer, you can also query the historical State of a thread at any time without re-executing the graph:

# 获取某个 thread 的当前（最新）State 快照
snapshot = react_graph_memory.get_state(config)
print(snapshot.values["messages"])   # 查看完整消息历史
print(snapshot.next)               # 查看下一步将执行哪个节点

# 获取某个 thread 的全部历史 checkpoints（时间旅行用）
for checkpoint in react_graph_memory.get_state_history(config):
    print(checkpoint.config)          # 该 checkpoint 的 config（含 checkpoint_id）
    print(checkpoint.values["messages"])  # 该时刻的消息列表

10Summary: What did you learn from this lecture?

This lecture is based on the complete ReAct Agent in 1-4, usingminimal code changesImplemented persistent multi-round dialogue memory.