Save and resume a local LLM workflow with LangGraph

LangGraph represents processing steps as nodes and their connections as edges. Saving state in a checkpoint lets the next request resume with the same thread_id. Start here with one question-answering node and a local Ollama call.

Start Ollama and note the chat model ID you already downloaded. In a Python virtual environment¹, install langgraph, langgraph-checkpoint-sqlite, and langchain-ollama. Set ChatOllama's base_url to the default server at http://127.0.0.1:11434 and use the model ID shown by ollama list.

python -m venv .venv
source .venv/bin/activate
python -m pip install -U langgraph langgraph-checkpoint-sqlite langchain-ollama

Use MessagesState and StateGraph to define one node that answers a user question. Open a SQLite saver as a context manager to store checkpoints in checkpoints.sqlite, then attach the saver when compiling the graph. Pass {'configurable': {'thread_id': 'demo-1'}} to the first and later invocations to continue that thread's messages.

import sys
from langchain_ollama import ChatOllama
from langgraph.checkpoint.sqlite import SqliteSaver
from langgraph.graph import START, END, MessagesState, StateGraph

model = ChatOllama(model="YOUR_OLLAMA_MODEL_ID", base_url="http://127.0.0.1:11434")
def answer(state: MessagesState):
    return {"messages": [model.invoke(state["messages"])]}

builder = StateGraph(MessagesState)
builder.add_node("answer", answer)
builder.add_edge(START, "answer")
builder.add_edge("answer", END)
config = {"configurable": {"thread_id": "demo-1"}}
mode = sys.argv[1]
with SqliteSaver.from_conn_string("checkpoints.sqlite") as checkpointer:
    graph = builder.compile(checkpointer=checkpointer)
    if mode == "start":
        question = "예제 문서: 장비 대여 규정의 시행일은 2026년 4월 1일이다. 이 문서의 시행일은 언제야?"
        graph.invoke({"messages": [("user", question)]}, config)
        print("saved messages:", len(graph.get_state(config).values["messages"]))
    elif mode == "resume":
        print("restored messages:", len(graph.get_state(config).values["messages"]))
        result = graph.invoke({"messages": [("user", "방금 답변에서 날짜만 다시 말해줘.")]}, config)
        print("messages after follow-up:", len(graph.get_state(config).values["messages"]))
        print(result["messages"][-1].content)
    else:
        raise SystemExit("Usage: python app.py start|resume")

Send an initial and follow-up question with the same thread_id, then inspect the length of graph.get_state(config).values['messages'] and the last reply. For a meaningful continuity check, do not manually resend the earlier conversation in the second question. After ending the process, reopen the same SQLite file with the same thread_id to load the saved state.

InMemorySaver suits an in-process check, but its state disappears when the process exits. If the SQLite file is not updated, check that the graph received the checkpointer and that each call uses the exact same thread_id. A different ID starts a separate conversation.

If several people continue the same conversation or you switch devices mid-thread, local checkpoints² alone do not solve it. Compare a setup that keeps state on a server separately.