Persisting Conversation History

Where we left off

Every lesson so far built contents fresh at the start of main() and threw it away when the script ended. A real assistant needs to survive restarts, a user closing a terminal and reopening it later, a server process being redeployed, without forgetting what was already asked. This lesson saves the entire structured transcript, questions, tool calls, and tool results alike, to disk, and reloads it into a working session.

The code, piece by piece

def save_history(contents: list[types.Content], path: Path) -> None:
    raw = [content.model_dump(mode="json") for content in contents]
    path.write_text(json.dumps(raw, indent=2))

google-genai's types build on pydantic.BaseModel, so every Content, and everything nested inside it (Part, FunctionCall, FunctionResponse), already knows how to serialize itself to plain JSON via model_dump(mode="json"). Nothing about function calls needed special-casing here, it's a direct consequence of Lesson 12's insight: contents is a structured object, and pydantic models serialize structure for free.

def load_history(path: Path) -> list[types.Content]:
    if not path.exists():
        return []
    raw = json.loads(path.read_text())
    return [types.Content.model_validate(item) for item in raw]

Content.model_validate() is the reverse direction, plain JSON back into real Content objects Gemini's SDK expects. An empty or missing file returns an empty history, a fresh session, exactly the shape a real assistant's first-ever run needs.

reloaded_contents = load_history(HISTORY_PATH)
answer = ask_turn(reloaded_contents, "What about it makes it smell like acetone?", store)

The demo deliberately simulates a restart: main() saves after turn one, then calls load_history() again as if starting a brand new process, and asks a follow-up question that only makes sense with turn one's context still present ("it" refers to the starter, established two questions ago).

Why this has to be the whole `contents` list, not just a Q&A log

A tempting shortcut is to persist only the final question/answer pairs, a plain list of {"question": ..., "answer": ...} dicts, since that's "the part a human cares about." That throws away exactly the structure Lesson 12 showed is load-bearing: which tool was called, with what arguments, and what it returned. Reloading only Q&A pairs and replaying them as plain text would mean re-deriving the model's own past reasoning from scratch every session, defeating the purpose.