短期记忆 - Docs by LangChain

概述

记忆是一个记住先前交互信息的系统。对于AI代理来说，记忆至关重要，因为它让代理能够记住先前的交互、从反馈中学习并适应用户偏好。随着代理处理更复杂的任务和大量的用户交互，这种能力对于效率和用户满意度都变得至关重要。短期记忆让你的应用程序能够在单个线程或对话中记住先前的交互。

一个线程组织会话中的多次交互，类似于电子邮件将消息分组到单个对话中的方式。

对话历史是短期记忆最常见的形式。长对话对当今的大语言模型（LLM）构成了挑战；完整的历史记录可能无法放入LLM的上下文窗口中，导致上下文丢失或错误。即使你的模型支持完整的上下文长度，大多数LLM在长上下文上的表现仍然不佳。它们会被过时或离题的内容“分心”，同时响应时间更慢，成本更高。聊天模型使用消息来接受上下文，其中包括指令（系统消息）和输入（人类消息）。在聊天应用程序中，消息在人类输入和模型响应之间交替，形成一个随时间增长的消息列表。由于上下文窗口有限，许多应用程序可以从使用技术来移除或“遗忘”过时信息中受益。

需要跨对话记住信息？使用长期记忆来存储和检索跨不同线程和会话的用户特定或应用程序级数据。

使用方法

要为代理添加短期记忆（线程级持久化），你需要在创建代理时指定一个checkpointer。

LangChain的代理将短期记忆作为代理状态的一部分进行管理。通过将这些存储在图的状态中，代理可以访问给定对话的完整上下文，同时保持不同线程之间的分离。状态使用检查点持久化到数据库（或内存），以便线程可以随时恢复。短期记忆在代理被调用或步骤（如工具调用）完成时更新，并且在每个步骤开始时读取状态。

import { createAgent } from "langchain";
import { MemorySaver } from "@langchain/langgraph";

const checkpointer = new MemorySaver();

const agent = createAgent({
    model: "claude-sonnet-4-6",
    tools: [],
    checkpointer,
});

await agent.invoke(
    { messages: [{ role: "user", content: "hi! i am Bob" }] },
    { configurable: { thread_id: "1" } }
);

在生产环境中

在生产环境中，使用由数据库支持的检查点：

import { PostgresSaver } from "@langchain/langgraph-checkpoint-postgres";

const DB_URI = "postgresql://postgres:postgres@localhost:5442/postgres?sslmode=disable";
const checkpointer = PostgresSaver.fromConnString(DB_URI);

有关更多检查点选项，包括SQLite、Postgres和Azure Cosmos DB，请参阅持久化文档中的检查点库列表。

自定义代理记忆

你可以通过创建带有状态模式的自定义中间件来扩展代理状态。自定义状态模式可以使用中间件中的stateSchema参数传递。最好使用StateSchema类进行状态定义（也支持普通的Zod对象）。

import { createAgent, createMiddleware } from "langchain";
import { StateSchema, MemorySaver } from "@langchain/langgraph";
import * as z from "zod";

const CustomState = new StateSchema({
    userId: z.string(),
    preferences: z.record(z.string(), z.any()),
});

const stateExtensionMiddleware = createMiddleware({
    name: "StateExtension",
    stateSchema: CustomState,
});

const checkpointer = new MemorySaver();
const agent = createAgent({
    model: "gpt-5.4",
    tools: [],
    middleware: [stateExtensionMiddleware],
    checkpointer,
});

// 自定义状态可以在调用时传递
const result = await agent.invoke({
    messages: [{ role: "user", content: "Hello" }],
    userId: "user_123",
    preferences: { theme: "dark" },
});

常见模式

启用短期记忆后，长对话可能会超出LLM的上下文窗口。常见的解决方案有：

裁剪消息

移除前N条或后N条消息（在调用LLM之前）

删除消息

从LangGraph状态中永久删除消息

总结消息

总结历史记录中较早的消息，并用摘要替换它们

自定义策略

自定义策略（例如，消息过滤等）

这允许代理跟踪对话而不会超出LLM的上下文窗口。

裁剪消息

大多数LLM都有一个最大支持的上下文窗口（以令牌为单位）。决定何时截断消息的一种方法是计算消息历史记录中的令牌数，并在接近该限制时进行截断。如果你使用LangChain，你可以使用裁剪消息实用程序，并指定要从列表中保留的令牌数量，以及用于处理边界的strategy（例如，保留最后maxTokens条）。要在代理中裁剪消息历史记录，请使用带有beforeModel钩子的createMiddleware：

import { RemoveMessage } from "@langchain/core/messages";
import { createAgent, createMiddleware } from "langchain";
import { MemorySaver, REMOVE_ALL_MESSAGES } from "@langchain/langgraph";

const trimMessages = createMiddleware({
  name: "TrimMessages",
  beforeModel: (state) => {
    const messages = state.messages;

    if (messages.length <= 3) {
      return; // 无需更改
    }

    const firstMsg = messages[0];
    const recentMessages =
      messages.length % 2 === 0 ? messages.slice(-3) : messages.slice(-4);
    const newMessages = [firstMsg, ...recentMessages];

    return {
      messages: [
        new RemoveMessage({ id: REMOVE_ALL_MESSAGES }),
        ...newMessages,
      ],
    };
  },
});

const checkpointer = new MemorySaver();
const agent = createAgent({
  model: "gpt-5.4",
  tools: [],
  middleware: [trimMessages],
  checkpointer,
});

删除消息

你可以从图状态中删除消息以管理消息历史记录。当你想要移除特定消息或清除整个消息历史记录时，这很有用。要从图状态中删除消息，你可以使用RemoveMessage。要使RemoveMessage工作，你需要使用带有messagesStateReducer reducer的状态键，例如MessagesValue。要移除特定消息：

import { RemoveMessage } from "@langchain/core/messages";

const deleteMessages = (state) => {
    const messages = state.messages;
    if (messages.length > 2) {
        // 移除最早的两条消息
        return {
        messages: messages
            .slice(0, 2)
            .map((m) => new RemoveMessage({ id: m.id })),
        };
    }
};

删除消息时，请确保生成的消息历史记录是有效的。检查你使用的LLM提供商的限制。例如：

一些提供商期望消息历史记录以user消息开头
大多数提供商要求带有工具调用的assistant消息后面必须跟有相应的tool结果消息。

import { RemoveMessage } from "@langchain/core/messages";
import { createAgent, createMiddleware } from "langchain";
import { MemorySaver } from "@langchain/langgraph";

const deleteOldMessages = createMiddleware({
  name: "DeleteOldMessages",
  afterModel: (state) => {
    const messages = state.messages;
    if (messages.length > 2) {
      // 移除最早的两条消息
      return {
        messages: messages
          .slice(0, 2)
          .map((m) => new RemoveMessage({ id: m.id! })),
      };
    }
    return;
  },
});

const agent = createAgent({
  model: "gpt-5.4",
  tools: [],
  systemPrompt: "Please be concise and to the point.",
  middleware: [deleteOldMessages],
  checkpointer: new MemorySaver(),
});

const config = { configurable: { thread_id: "1" } };

const streamA = await agent.stream(
  { messages: [{ role: "user", content: "hi! I'm bob" }] },
  { ...config, streamMode: "values" }
);
for await (const event of streamA) {
  const messageDetails = event.messages.map((message) => [
    message.getType(),
    message.content,
  ]);
  console.log(messageDetails);
}

const streamB = await agent.stream(
  {
    messages: [{ role: "user", content: "what's my name?" }],
  },
  { ...config, streamMode: "values" }
);
for await (const event of streamB) {
  const messageDetails = event.messages.map((message) => [
    message.getType(),
    message.content,
  ]);
  console.log(messageDetails);
}

[[ "human", "hi! I'm bob" ]]
[[ "human", "hi! I'm bob" ], [ "ai", "Hello, Bob! How can I assist you today?" ]]
[[ "human", "hi! I'm bob" ], [ "ai", "Hello, Bob! How can I assist you today?" ]]
[[ "human", "hi! I'm bob" ], [ "ai", "Hello, Bob! How can I assist you today" ], ["human", "what's my name?" ]]
[[ "human", "hi! I'm bob" ], [ "ai", "Hello, Bob! How can I assist you today?" ], ["human", "what's my name?"], [ "ai", "Your name is Bob, as you mentioned. How can I help you further?" ]]
[[ "human", "what's my name?" ], [ "ai", "Your name is Bob, as you mentioned. How can I help you further?" ]]

总结消息

如上所示，裁剪或删除消息的问题在于，你可能会因消息队列的筛选而丢失信息。因此，一些应用程序受益于使用聊天模型总结消息历史记录的更复杂方法。

要在代理中总结消息历史记录，请使用内置的summarizationMiddleware：

import { createAgent, summarizationMiddleware } from "langchain";
import { MemorySaver } from "@langchain/langgraph";

const checkpointer = new MemorySaver();

const agent = createAgent({
  model: "gpt-5.4",
  tools: [],
  middleware: [
    summarizationMiddleware({
      model: "gpt-5.4-mini",
      trigger: { tokens: 4000 },
      keep: { messages: 20 },
    }),
  ],
  checkpointer,
});

const config = { configurable: { thread_id: "1" } };
await agent.invoke({ messages: "hi, my name is bob" }, config);
await agent.invoke({ messages: "write a short poem about cats" }, config);
await agent.invoke({ messages: "now do the same but for dogs" }, config);
const finalResponse = await agent.invoke({ messages: "what's my name?" }, config);

console.log(finalResponse.messages.at(-1)?.content);
// Your name is Bob!

有关更多配置选项，请参阅summarizationMiddleware。

访问记忆

你可以通过多种方式访问和修改代理的短期记忆（状态）：

工具

在工具中读取短期记忆

使用runtime参数（类型为ToolRuntime）在工具中访问短期记忆（状态）。 runtime参数在工具签名中是隐藏的（因此模型看不到它），但工具可以通过它访问状态。

import { createAgent, tool, type ToolRuntime } from "langchain";
import { StateSchema } from "@langchain/langgraph";
import * as z from "zod";

const CustomState = new StateSchema({
  userId: z.string(),
});

const getUserInfo = tool(
  async (_, config: ToolRuntime<typeof CustomState.State>) => {
    const userId = config.state.userId;
    return userId === "user_123" ? "John Doe" : "Unknown User";
  },
  {
    name: "get_user_info",
    description: "Get user info",
    schema: z.object({}),
  }
);

const agent = createAgent({
  model: "gpt-5-nano",
  tools: [getUserInfo],
  stateSchema: CustomState,
});

const result = await agent.invoke(
  {
    messages: [{ role: "user", content: "what's my name?" }],
    userId: "user_123",
  },
  {
    context: {},
  }
);

console.log(result.messages.at(-1)?.content);
// 输出: "Your name is John Doe."

从工具写入短期记忆

要在执行期间修改代理的短期记忆（状态），你可以直接从工具返回状态更新。这对于持久化中间结果或使信息可供后续工具或提示访问很有用。

import { tool, createAgent, ToolMessage, type ToolRuntime } from "langchain";
import { Command, StateSchema } from "@langchain/langgraph";
import * as z from "zod";

const CustomState = new StateSchema({
  userId: z.string().optional(),
});

const updateUserInfo = tool(
  async (_, config: ToolRuntime<typeof CustomState.State>) => {
    const userId = config.state.userId;
    const name = userId === "user_123" ? "John Smith" : "Unknown user";
    return new Command({
      update: {
        userName: name,
        // 更新消息历史记录
        messages: [
          new ToolMessage({
            content: "Successfully looked up user information",
            tool_call_id: config.toolCall?.id ?? "",
          }),
        ],
      },
    });
  },
  {
    name: "update_user_info",
    description: "Look up and update user info.",
    schema: z.object({}),
  }
);

const greet = tool(
  async (_, config) => {
    const userName = config.context?.userName;
    return `Hello ${userName}!`;
  },
  {
    name: "greet",
    description: "Use this to greet the user once you found their info.",
    schema: z.object({}),
  }
);

const agent = createAgent({
  model: "openai:gpt-5-mini",
  tools: [updateUserInfo, greet],
  stateSchema: CustomState,
});

const result = await agent.invoke({
  messages: [{ role: "user", content: "greet the user" }],
  userId: "user_123",
});

console.log(result.messages.at(-1)?.content);
// 输出: "Hello! I’m here to help — what would you like to do today?"

提示

在中间件中访问短期记忆（状态），以根据对话历史记录或自定义状态字段创建动态提示。

import * as z from "zod";
import { createAgent, tool, dynamicSystemPromptMiddleware } from "langchain";

const contextSchema = z.object({
  userName: z.string(),
});
type ContextSchema = z.infer<typeof contextSchema>;

const getWeather = tool(
  async ({ city }) => {
    return `The weather in ${city} is always sunny!`;
  },
  {
    name: "get_weather",
    description: "Get user info",
    schema: z.object({
      city: z.string(),
    }),
  }
);

const agent = createAgent({
  model: "gpt-5-nano",
  tools: [getWeather],
  contextSchema,
  middleware: [
    dynamicSystemPromptMiddleware<ContextSchema>((_, config) => {
      return `You are a helpful assistant. Address the user as ${config.context?.userName}.`;
    }),
  ],
});

const result = await agent.invoke(
  {
    messages: [{ role: "user", content: "What is the weather in SF?" }],
  },
  {
    context: {
      userName: "John Smith",
    },
  }
);

for (const message of result.messages) {
  console.log(message);
}
/**
 * HumanMessage {
 *   "content": "What is the weather in SF?",
 *   // ...
 * }
 * AIMessage {
 *   // ...
 *   "tool_calls": [
 *     {
 *       "name": "get_weather",
 *       "args": {
 *         "city": "San Francisco"
 *       },
 *       "type": "tool_call",
 *       "id": "call_tCidbv0apTpQpEWb3O2zQ4Yx"
 *     }
 *   ],
 *   // ...
 * }
 * ToolMessage {
 *   "content": "The weather in San Francisco is always sunny!",
 *   "tool_call_id": "call_tCidbv0apTpQpEWb3O2zQ4Yx"
 *   // ...
 * }
 * AIMessage {
 *   "content": "John Smith, here's the latest: The weather in San Francisco is always sunny!\n\nIf you'd like more details (temperature, wind, humidity) or a forecast for the next few days, I can pull that up. What would you like?",
 *   // ...
 * }
 */

模型之前

import { RemoveMessage } from "@langchain/core/messages";
import { createAgent, createMiddleware, trimMessages } from "langchain";
import { MemorySaver } from "@langchain/langgraph";
import { REMOVE_ALL_MESSAGES } from "@langchain/langgraph";

const trimMessageHistory = createMiddleware({
  name: "TrimMessages",
  beforeModel: async (state) => {
    const trimmed = await trimMessages(state.messages, {
      maxTokens: 384,
      strategy: "last",
      startOn: "human",
      endOn: ["human", "tool"],
      tokenCounter: (msgs) => msgs.length,
    });
    return {
      messages: [new RemoveMessage({ id: REMOVE_ALL_MESSAGES }), ...trimmed],
    };
  },
});

const checkpointer = new MemorySaver();
const agent = createAgent({
  model: "gpt-5-nano",
  tools: [],
  middleware: [trimMessageHistory],
  checkpointer,
});

模型之后

import { RemoveMessage } from "@langchain/core/messages";
import { createAgent, createMiddleware } from "langchain";
import { REMOVE_ALL_MESSAGES } from "@langchain/langgraph";

const validateResponse = createMiddleware({
  name: "ValidateResponse",
  afterModel: (state) => {
    const lastMessage = state.messages.at(-1)?.content;
    if (
      typeof lastMessage === "string" &&
      lastMessage.toLowerCase().includes("confidential")
    ) {
      return {
        messages: [
          new RemoveMessage({ id: REMOVE_ALL_MESSAGES }),
        ],
      };
    }
    return;
  },
});

const agent = createAgent({
  model: "gpt-5-nano",
  tools: [],
  middleware: [validateResponse],
});

将这些文档连接到Claude、VSCode等，通过MCP获取实时答案。

在GitHub上编辑此页面或提交问题。

​概述

​使用方法

​在生产环境中

​自定义代理记忆

​常见模式

裁剪消息

删除消息

总结消息

自定义策略

​裁剪消息

​删除消息

​总结消息

​访问记忆

​工具

​在工具中读取短期记忆

​从工具写入短期记忆

​提示

​模型之前

​模型之后

概述

使用方法

在生产环境中

自定义代理记忆

常见模式

裁剪消息

删除消息

总结消息

访问记忆

工具

在工具中读取短期记忆

从工具写入短期记忆

提示

模型之前

模型之后