Persistent Checkpointer

The limitation we've carried since Lesson 24

InMemorySaver (Lesson 24) gives an agent real memory across turns, but that memory lives only inside the running Python process. Close the program, and it's gone, the exact same limitation Lesson 17 flagged for a plain Python list. This lesson swaps InMemorySaver for SqliteSaver, which writes to a real file on disk instead.

Swapping the checkpointer, nothing else changes

with SqliteSaver.from_conn_string(str(DB_PATH)) as checkpointer:
    agent = create_agent(model=model, checkpointer=checkpointer)

Compare this to Lesson 24's checkpointer=InMemorySaver(). The rest of create_agent, the rest of how you call .invoke(), the thread_id pattern, none of it changes. This is exactly the payoff promised back in Lesson 24: swapping to a persistent checkpointer is a one-line change, because everything else was already built against the same checkpointer= interface.

SqliteSaver.from_conn_string(...) opens a connection to a real .db file (SQLite is a lightweight, file-based database, no separate server needed) and returns it as a context manager, hence the with block, the connection gets cleanly closed when the block ends.

Simulating two separate program runs

def session_one() -> None:
    with SqliteSaver.from_conn_string(str(DB_PATH)) as checkpointer:
        agent = create_agent(model=model, checkpointer=checkpointer)
        ...

def session_two() -> None:
    with SqliteSaver.from_conn_string(str(DB_PATH)) as checkpointer:
        agent = create_agent(model=model, checkpointer=checkpointer)
        ...

session_two() builds a completely new agent object, with a completely new connection, sharing nothing with session_one() except the thread_id and the path to the same .db file. This is deliberately structured this way to prove a point: if session_two() still remembers what was said in session_one(), it can't be because some Python object was kept around in memory, there's no shared object at all, only a shared file on disk.

Proof it worked

Run the lesson: session one tells the agent "My favorite programming language is Python." Session two, a genuinely separate agent and connection, is asked "What is my favorite programming language?" and answers correctly: "Your favorite programming language is Python!" The only place that fact could have come from is the memory.db file itself.

print(f"File size: {DB_PATH.stat().st_size} bytes")

That memory.db file is real, sitting on disk in this lesson's folder after you run it, you can inspect it, back it up, or delete it like any other file (which is exactly why it's excluded from this project's git history, generated data, not source code).