先看存储接口 BaseChatMessageHistory（libs/core/langchain_core/chat_history.py:22）。接口非常小：一个"读"（messages 属性）、一个"写"（add_messages）、一个"清空"（clear）：

# libs/core/langchain_core/chat_history.py:169
def add_messages(self, messages: Sequence[BaseMessage]) -> None:
    """Add a list of messages. ..."""
    for message in messages:
        self.add_message(message)          # 默认逐条写；子类可覆写成"批量写"更高效

# libs/core/langchain_core/chat_history.py:202
class InMemoryChatMessageHistory(BaseChatMessageHistory, BaseModel):
    """In memory implementation of chat message history."""

    messages: list[BaseMessage] = Field(default_factory=list)   # ★存储就是一个 list

    def add_message(self, message: BaseMessage) -> None:
        self.messages.append(message)      # 写 = append

    def clear(self) -> None:
        self.messages = []                 # 清空 = 换个空 list

主角登场：RunnableWithMessageHistory（libs/core/langchain_core/runnables/history.py:39）。它的 __init__（history.py:250）把你的链组装成一条"带前后钩子"的新链：

# libs/core/langchain_core/runnables/history.py:332（__init__ 内部，节选）
history_chain: Runnable[Any, Any] = RunnableLambda(
    self._enter_history, self._aenter_history          # ① 前置步骤：进场读历史
).with_config(run_name="load_history")

messages_key = history_messages_key or input_messages_key
if messages_key:
    history_chain = RunnablePassthrough.assign(        # ② 读到的历史塞进输入 dict 的指定 key
        **{messages_key: history_chain}
    ).with_config(run_name="insert_history")

runnable_sync = runnable.with_listeners(on_end=self._exit_history)   # ③ 后置钩子：结束时写历史
runnable_async = runnable.with_alisteners(on_end=self._aexit_history)

bound = (
    history_chain                                       # ④ 新链 = 读历史 | 你的链(带写历史钩子)
    | RunnableLambda(_call_runnable_sync, _call_runnable_async)
      .with_config(run_name="check_sync_or_async")
).with_config(run_name="RunnableWithMessageHistory")

调用时你传 config={"configurable": {"session_id": "foo"}}，护士怎么据此取到"foo 的病历"？答案在 _merge_configs（libs/core/langchain_core/runnables/history.py:581）——它在每次调用合并 config 时被触发：

# libs/core/langchain_core/runnables/history.py:581
def _merge_configs(self, *configs: RunnableConfig | None) -> RunnableConfig:
    config = super()._merge_configs(*configs)
    expected_keys = [field_spec.id for field_spec in self.history_factory_config]
    configurable = config.get("configurable", {})

    missing_keys = set(expected_keys) - set(configurable.keys())
    if missing_keys and parameter_names:
        msg = (f"Missing keys {sorted(missing_keys)} in config['configurable'] ...")
        raise ValueError(msg)                              # 没带 session_id？直接报错教你怎么传

    if len(expected_keys) == 1:
        message_history = self.get_session_history(        # ★单 key：位置参数调工厂
            configurable[expected_keys[0]]                 #   get_session_history("foo")
        )
    else:
        message_history = self.get_session_history(        # 多 key：按名字传
            **{key: configurable[key] for key in expected_keys}  # (user_id=..., conversation_id=...)
        )
    config["configurable"]["message_history"] = message_history  # ★取到的"病历本"塞回 config
    return config

最后看两个钩子本体。进场 _enter_history（libs/core/langchain_core/runnables/history.py:521）+ 退场 _exit_history（history.py:547）：

# libs/core/langchain_core/runnables/history.py:521
def _enter_history(self, value: Any, config: RunnableConfig) -> list[BaseMessage]:
    hist: BaseChatMessageHistory = config["configurable"]["message_history"]  # 从 config 摸出病历本
    messages = hist.messages.copy()                       # ① 翻出全部历史
    if not self.history_messages_key:
        input_val = (value if not self.input_messages_key
                     else value[self.input_messages_key])
        messages += self._get_input_messages(input_val)   # ② 没单独 history key 时，历史+本轮输入拼一起
    return messages

# libs/core/langchain_core/runnables/history.py:547
def _exit_history(self, run: Run, config: RunnableConfig) -> None:
    hist: BaseChatMessageHistory = config["configurable"]["message_history"]
    inputs = load(run.inputs, allowed_objects="messages")
    input_messages = self._get_input_messages(inputs)     # ③ 从 Run 记录还原本轮输入
    if not self.history_messages_key:
        historic_messages = config["configurable"]["message_history"].messages
        input_messages = input_messages[len(historic_messages):]  # ④ ★去掉开头的历史，防止重复入库
    output_val = load(run.outputs, allowed_objects="messages")
    output_messages = self._get_output_messages(output_val)
    hist.add_messages(input_messages + output_messages)   # ⑤ 本轮"问+答"批量补写进病历

读 __init__ 时你可能注意到开头有段醒目的代码（libs/core/langchain_core/runnables/history.py:324）：

# libs/core/langchain_core/runnables/history.py:324
warn_deprecated(
    since="1.3.3",
    message=(
        "RunnableWithMessageHistory is deprecated. "
        "Use LangGraph's built-in persistence instead."   # 官方指路：用 LangGraph 的持久化
    ),
    removal="2.0.0",
)

cd /Users/bitmart/work/codes/github/AI_WORK/langchain/libs/core/langchain_core

# 1. 记事本抽象：读/写/清空 + 内存实现
sed -n '22,50p'   chat_history.py       # BaseChatMessageHistory 与实现指南
sed -n '202,247p' chat_history.py       # InMemoryChatMessageHistory

# 2. 护士的组装与钩子
sed -n '324,357p' runnables/history.py  # 废弃警告 + history_chain 组装
sed -n '521,562p' runnables/history.py  # _enter_history / _exit_history
sed -n '581,626p' runnables/history.py  # _merge_configs：session_id → 病历本

# 3. 输入归一化的宽容度
sed -n '452,489p' runnables/history.py  # _get_input_messages