Checkpointer Memory

Where we left off

Lesson 6 put a model inside a node. Lesson 7 gave that node tools. But every .invoke() in the beginner tier was still independent, whatever messages you passed in were the only ones the graph ever saw. Ask a follow-up question in a second call, and the graph has no idea a first call ever happened. This lesson fixes that with a checkpointer.

The checkpointer: automatic memory, not a list you manage

app = builder.compile(checkpointer=InMemorySaver())

That's the only structural change from a Lesson 6 style graph. A checkpointer saves a snapshot of the graph's state after every node finishes running, and it saves each snapshot under a thread ID, so many separate conversations can share one compiled graph without crosstalk. InMemorySaver is the simplest kind: snapshots live in this process's RAM, and are gone the instant the program exits (Lesson 14 swaps this for a version that survives a restart).

thread_id: which conversation is this?

config_a = {"configurable": {"thread_id": "conversation-a"}}

result1 = app.invoke({"messages": [HumanMessage("My favorite color is teal.")]}, config_a)
result2 = app.invoke({"messages": [HumanMessage("What is my favorite color?")]}, config_a)

Both calls pass config_a, same thread_id. Before chatbot runs on the second call, the checkpointer looks up everything saved earlier under "conversation-a" and prepends it to state, so the model sees the full conversation, not just the new question. Notice each call only sends the new message, exactly like create_agent plus InMemorySaver in the LangChain course, except here there's no create_agent doing it for us, the checkpointer is doing this directly against our own hand-built graph.

A different thread_id is a different conversation

config_b = {"configurable": {"thread_id": "conversation-b"}}
result3 = app.invoke({"messages": [HumanMessage("What is my favorite color?")]}, config_b)

Same compiled app, same process, same everything except the thread_id. The model correctly has no idea what color was mentioned, because "conversation-b" has no saved history yet. This is how one compiled graph serves many independent users or sessions at once, memory scoping is entirely a function of thread_id, not of which Python object you're holding.