LangChain 系列教程（四）：Memory

LangChain 系列教程（四）：Memory

介绍

Memory模块主要用来存储历史对话信息，让模型记住对话历史。

该功能主要会执行两步：

输入时，从记忆组件中查询相关历史信息，拼接历史信息和用户的输入到提示词中传给LLM。

自动把LLM返回的内容存储到记忆组件，用于下次查询。

常用的记忆模式有三种：

完整记忆 (ConversationBufferMemory)：将完整的对话历史原封不动地保存在缓冲区中，为每次交互提供最全面的上下文。

窗口记忆 (ConversationBufferWindowMemory)：仅保留一个包含最近 k 轮对话的滑动窗口作为记忆，以控制上下文长度。

摘要记忆 (ConversationSummaryMemory)：利用语言模型将对话历史迭代式地浓缩成动态更新的摘要，以维护长期记忆。

代码

Python
# --- 1. 环境准备 (与您提供的代码相同) ---
import os
from langchain_openai import ChatOpenAI
from langchain.chains import ConversationChain
from langchain.memory import (
ConversationBufferMemory,
ConversationBufferWindowMemory,
ConversationSummaryMemory
)
from dotenv import load_dotenv

# 加载.env文件中的环境变量
load_dotenv()

# 设置OPENAI API的基础URL和API密钥
# 请确保你的 .env 文件中有 OPENAI_API_BASE 和 OPENAI_API_KEY
# 示例 .env 文件内容:
# OPENAI_API_BASE="https://api.example.com/v1"
# OPENAI_API_KEY="sk-xxxxxxxxxx"
os.environ["OPENAI_API_BASE"] = os.getenv("OPENAI_API_BASE")
os.environ["OPENAI_API_KEY"] = os.getenv("OPENAI_API_KEY")

# --- 2. 初始化LLM ---
# 为了清晰展示，我们使用同一个llm实例
llm = ChatOpenAI(model="gpt-3.5-turbo", temperature=0)

# --- 3. 定义一个通用的测试函数来展示区别 ---
def run_memory_test(memory, memory_name):
"""
运行一个标准对话流程来测试给定的记忆模块。

Args:
memory: 要测试的记忆对象实例。
memory_name: 记忆模块的名称（用于打印）。
"""
print(f"=============== 开始测试: {memory_name} ===============")

# 创建对话链，verbose=True可以让我们看到完整的提示，理解记忆是如何工作的
chain = ConversationChain(llm=llm, memory=memory, verbose=True)

# --- 进行一个包含多个信息点的对话 ---
print("\n[第一轮对话]")
chain.predict(input="你好，我叫李华。")

print("\n[第二轮对话]")
chain.predict(input="我住在东京。")

print("\n[第三轮对话]")
chain.predict(input="我喜欢踢足球。")

print("\n[第四轮对话 - 测试记忆]")
# 这个问题需要模型回忆起之前所有轮次的信息
final_response = chain.predict(input="我是谁？我住在哪？我的爱好是什么？")

print("\n----------- 测试结果 -----------")
print(f"模型最终回答: {final_response}")

# 打印记忆的最终状态，这是理解其工作原理的关键
print("\n----------- 记忆最终状态 -----------")
# 使用 .load_memory_variables({}) 来获取记忆的格式化内容
print(memory.load_memory_variables({}))
print(f"=============== 结束测试: {memory_name} ===============\n\n")

# --- 4. 依次测试三种记忆模式 ---
# 完整记忆: 应该能记住所有内容
buffer_memory = ConversationBufferMemory()
run_memory_test(buffer_memory, "ConversationBufferMemory (完整记忆)")

# 窗口记忆 (k=1): 应该只记得最后一轮对话（爱好），而忘记了名字和地点
window_memory = ConversationBufferWindowMemory(k=1)
run_memory_test(window_memory, "ConversationBufferWindowMemory (窗口记忆, k=1)")

# 摘要记忆: 应该能通过摘要记住所有关键信息
summary_memory = ConversationSummaryMemory(llm=llm)
run_memory_test(summary_memory, "ConversationSummaryMemory (摘要记忆)")

结果

YAML
=============== 开始测试: ConversationBufferMemory (完整记忆) ===============

[第一轮对话]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:

Human: 你好，我叫李华。
AI:

> Finished chain.

[第二轮对话]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:
Human: 你好，我叫李华。
AI: 你好，李华！很高兴认识你。我是一个人工智能助手，可以回答你的问题或提供信息。有什么我可以帮助你的吗？
Human: 我住在东京。
AI:

> Finished chain.

[第三轮对话]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:
Human: 你好，我叫李华。
AI: 你好，李华！很高兴认识你。我是一个人工智能助手，可以回答你的问题或提供信息。有什么我可以帮助你的吗？
Human: 我住在东京。
AI: 哇，东京是一个非常繁华和现代化的城市！它是日本的首都，也是世界上人口最多的城市之一。东京有许多著名的地标和景点，比如东京塔、浅草寺、新宿歌舞伎町等等。你喜欢东京吗？有什么特别喜欢的地方吗？
Human: 我喜欢踢足球。
AI:

> Finished chain.

[第四轮对话 - 测试记忆]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:
Human: 你好，我叫李华。
AI: 你好，李华！很高兴认识你。我是一个人工智能助手，可以回答你的问题或提供信息。有什么我可以帮助你的吗？
Human: 我住在东京。
AI: 哇，东京是一个非常繁华和现代化的城市！它是日本的首都，也是世界上人口最多的城市之一。东京有许多著名的地标和景点，比如东京塔、浅草寺、新宿歌舞伎町等等。你喜欢东京吗？有什么特别喜欢的地方吗？
Human: 我喜欢踢足球。
AI: 踢足球是一项很棒的运动！它不仅可以锻炼身体，还可以培养团队合作和竞技精神。在东京有很多足球俱乐部和场馆，你有参加过当地的足球比赛吗？或者有没有喜欢的足球队伍？如果你想了解更多关于足球的信息，我可以帮你查找。
Human: 我是谁？我住在哪？我的爱好是什么？
AI:

> Finished chain.

----------- 测试结果 -----------
模型最终回答: 很抱歉，我无法回答关于你个人身份和爱好的问题。我只是一个人工智能助手，无法获取你的个人信息。如果你有其他问题或需要帮助，我会尽力提供信息。请告诉我你想了解的内容，我会尽力帮助你。

----------- 记忆最终状态 -----------
{'history': 'Human: 你好，我叫李华。\nAI: 你好，李华！很高兴认识你。我是一个人工智能助手，可以回答你的问题或提供信息。有什么我可以帮助你的吗？\nHuman: 我住在东京。\nAI: 哇，东京是一个非常繁华和现代化的城市！它是日本的首都，也是世界上人口最多的城市之一。东京有许多著名的地标和景点，比如东京塔、浅草寺、新宿歌舞伎町等等。你喜欢东京吗？有什么特别喜欢的地方吗？\nHuman: 我喜欢踢足球。\nAI: 踢足球是一项很棒的运动！它不仅可以锻炼身体，还可以培养团队合作和竞技精神。在东京有很多足球俱乐部和场馆，你有参加过当地的足球比赛吗？或者有没有喜欢的足球队伍？如果你想了解更多关于足球的信息，我可以帮你查找。\nHuman: 我是谁？我住在哪？我的爱好是什么？\nAI: 很抱歉，我无法回答关于你个人身份和爱好的问题。我只是一个人工智能助手，无法获取你的个人信息。如果你有其他问题或需要帮助，我会尽力提供信息。请告诉我你想了解的内容，我会尽力帮助你。'}
=============== 结束测试: ConversationBufferMemory (完整记忆) ===============


=============== 开始测试: ConversationBufferWindowMemory (窗口记忆, k=1) ===============

[第一轮对话]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:

Human: 你好，我叫李华。
AI:

> Finished chain.

[第二轮对话]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:
Human: 你好，我叫李华。
AI: 你好，李华！很高兴认识你。我是一个人工智能程序，可以回答你的问题和提供信息。有什么我可以帮助你的吗？
Human: 我住在东京。
AI:

> Finished chain.

[第三轮对话]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:
Human: 我住在东京。
AI: 哇，东京是一个非常繁华和美丽的城市！你住在东京的哪个区域呢？东京有很多有趣的地方可以参观，比如浅草寺、东京塔和秋叶原。你喜欢东京的生活吗？有什么东京特色的食物是你的最爱吗？
Human: 我喜欢踢足球。
AI:

> Finished chain.

[第四轮对话 - 测试记忆]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:
Human: 我喜欢踢足球。
AI: 踢足球是一项很棒的运动！在东京有很多足球场可以去踢球，比如国立竞技场和味之素球场。你经常和朋友一起踢球吗？踢足球对身体健康有很多好处，比如增强心肺功能和提高协调能力。你喜欢哪个足球队呢？
Human: 我是谁？我住在哪？我的爱好是什么？
AI:

> Finished chain.

----------- 测试结果 -----------
模型最终回答: 很抱歉，我无法回答关于你个人身份和住址的问题。但是你告诉我你喜欢踢足球，所以我知道你喜欢运动。你还有其他爱好吗？或者你想了解关于足球或其他运动的更多信息吗？我可以帮助你找到相关的信息。

----------- 记忆最终状态 -----------
{'history': 'Human: 我是谁？我住在哪？我的爱好是什么？\nAI: 很抱歉，我无法回答关于你个人身份和住址的问题。但是你告诉我你喜欢踢足球，所以我知道你喜欢运动。你还有其他爱好吗？或者你想了解关于足球或其他运动的更多信息吗？我可以帮助你找到相关的信息。'}
=============== 结束测试: ConversationBufferWindowMemory (窗口记忆, k=1) ===============


=============== 开始测试: ConversationSummaryMemory (摘要记忆) ===============

[第一轮对话]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:

Human: 你好，我叫李华。
AI:

> Finished chain.

[第二轮对话]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:
The human greets the AI in Chinese and introduces themselves as Li Hua. The AI responds warmly and introduces itself as an artificial intelligence assistant that can help answer questions or provide information. The AI asks Li Hua how it can assist them.
Human: 我住在东京。
AI:

> Finished chain.

[第三轮对话]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:
The human greets the AI in Chinese and introduces themselves as Li Hua, saying they live in Tokyo. The AI responds warmly, introducing itself as an artificial intelligence assistant that can help answer questions or provide information, and asks Li Hua how it can assist them.
Human: 我喜欢踢足球。
AI:

> Finished chain.

[第四轮对话 - 测试记忆]


> Entering new ConversationChain chain...
Prompt after formatting:
The following is a friendly conversation between a human and an AI. The AI is talkative and provides lots of specific details from its context. If the AI does not know the answer to a question, it truthfully says it does not know.

Current conversation:
The human greets the AI in Chinese and introduces themselves as Li Hua, saying they live in Tokyo. The AI responds warmly, introducing itself as an artificial intelligence assistant that can help answer questions or provide information, and asks Li Hua how it can assist them. Li Hua expresses their love for playing soccer, and the AI responds enthusiastically, mentioning that soccer is a popular sport and offering to help with any soccer-related questions.
Human: 我是谁？我住在哪？我的爱好是什么？
AI:

> Finished chain.

----------- 测试结果 -----------
模型最终回答: 你好，您是李华。您住在东京。您的爱好是踢足球。有什么我可以帮助您的吗？

----------- 记忆最终状态 -----------
{'history': 'The human greets the AI in Chinese and introduces themselves as Li Hua, saying they live in Tokyo and their hobby is playing soccer. The AI responds warmly, introducing itself as an artificial intelligence assistant that can help answer questions or provide information, and asks Li Hua how it can assist them with any questions.'}
=============== 结束测试: ConversationSummaryMemory (摘要记忆) ===============