Three prefixes, three lifetimes

A key with no prefix belongs to this conversation. Three prefixes change that, and one of them is how an agent remembers you across conversations.

A tool that writes two of them

def remember(name: str, tool_context: ToolContext) -> dict:
    """Remember the customer's name for every future conversation."""
    tool_context.state["user:name"] = name
    tool_context.state["temp:working"] = "scratch"
    return {"status": "success"}

One key with the user prefix, one temporary. Everything else about the tool is the same as lesson 21.

agent = LlmAgent(
    name="support",
    model=PretendModel(replies=[call("remember", name="Asha"), say("Nice to meet you.")]),
    instruction="Be friendly.",
    tools=[remember],
)

runner = InMemoryRunner(agent=agent, app_name="demo")

Watching a prefix cross a session

first = await runner.session_service.create_session(app_name="demo", user_id="u1")
message = types.Content(role="user", parts=[types.Part(text="I am Asha")])

async for _ in runner.run_async(user_id="u1", session_id=first.id, new_message=message):
    pass

done = await runner.session_service.get_session(
    app_name="demo", user_id="u1", session_id=first.id)
print("first session: ", dict(done.state))

The user key is there, and temp:working is not. Temporary means the current invocation only, and that one has finished.

second = await runner.session_service.create_session(app_name="demo", user_id="u1")
print("second session:", dict(second.state))

other = await runner.session_service.create_session(app_name="demo", user_id="u2")
print("another user:  ", dict(other.state))

A brand new conversation for the same person already knows the name. A different person does not. That is the whole point of the prefix.

Where the persistence actually comes from

The prefixes describe scope. Whether anything survives a restart is a property of the session service. InMemorySessionService keeps user and app state across sessions, as you just saw, and loses everything when the process ends.

• Learning and tests: in memory, which is what InMemoryRunner gives you.
• A real deployment: a database session service, so a restart does not wipe conversations.
• On Google Cloud: the managed service, which persists for you.