Using InMemorySaver to Give LangGraph Agents Persistent Conversation Memory

Checkpoint: Core of State Persistence

LangGraph’s checkpoint mechanism records a snapshot after each agent step, allowing automatic restoration of context on the next turn.

Two core operations are used: checkpointer.put(config, checkpoint) – saves the snapshot. checkpointer.get(config) – retrieves it.

InMemorySaver: In‑memory Checkpoint for Development

In development environments the InMemorySaver class stores checkpoints in RAM; production uses PostgresSaver.

from langgraph.checkpoint.memory import InMemorySaver
from langgraph.graph import StateGraph, START, END
# ... define ConversationState, chat_node ...
builder = StateGraph(ConversationState)
builder.add_node("chat", chat_node)
builder.add_edge(START, "chat")
builder.add_edge("chat", END)

memory = InMemorySaver()
graph = builder.compile(checkpointer=memory)

thread_id: Isolating Separate Conversations

The configurable thread_id key distinguishes independent sessions, e.g., “user‑123” vs “user‑456”.

# Session A
config_a = {"configurable": {"thread_id": "user-123"}}
# Session B
config_b = {"configurable": {"thread_id": "user-456"}}
result_a = graph.invoke({"messages": ["你好"], "counter": 0}, config_a)
result_b = graph.invoke({"messages": ["你好"], "counter": 0}, config_b)

Manual State Control with get_state / update_state

Developers can fetch the current snapshot via graph.get_state(config) and modify fields directly with graph.update_state(config, {...}).

snapshot = graph.get_state(config_a)
print(snapshot.values)  # {'messages': ['你好', '回复: 你好'], 'counter': 1}
graph.update_state(config_a, {"counter": 100})

Time‑Travel Replay

Checkpoints retain execution history, enabling replay of any point by iterating over memory.list(config) and re‑executing.

for checkpoint in memory.list(config_a):
    print(f"ID: {checkpoint.id}")
    print(f"Time: {checkpoint.metadata}")
    print(f"Value: {checkpoint.values}")

Practical Use Cases

Recovering from errors by restoring a checkpoint.

Debugging specific steps via replay.

Implementing undo operations.

Full Example: A Conversational Agent with Memory

from langgraph.checkpoint.memory import InMemorySaver
from langgraph.prebuilt import create_react_agent

@tool
def multiply(a: int, b: int) -> int:
    """Multiply two numbers."""
    return a * b

model = ChatAnthropic(model="claude-sonnet-4-6")
memory = InMemorySaver()
agent = create_react_agent(model, [multiply], checkpointer=memory)

config = {"configurable": {"thread_id": "conversation-1"}}
result1 = agent.invoke({"messages": [{"role": "user", "content": "3乘4等于多少？"}]}, config)
print(result1["messages"][-1].content)  # "3乘4等于12"

result2 = agent.invoke({"messages": [{"role": "user", "content": "再加5呢？"}]}, config)
print(result2["messages"][-1].content)

Production with PostgresSaver

from langgraph.checkpoint.postgres import PostgresSaver

saver = PostgresSaver.from_conn_string(
    "postgresql://user:password@localhost:5432/langgraph"
)
saver.setup()  # run once to create tables
graph = builder.compile(checkpointer=saver)