feat(anthropic): implement Anthropic memory tool in provider package

[christian-bromann]

This PR adds support for Anthropic's new Memory Tool API to @langchain/anthropic, enabling Claude to store and retrieve information across conversations through a persistent memory file directory.

What's New

Memory Tool (memory_20250818)

A new tool that allows Claude to manage persistent memory through file-based operations:

• view - List directory contents or read file content
• create - Create new memory files
• str_replace - Update file content by replacing text
• insert - Insert text at specific line numbers
• delete - Remove memory files
• rename - Rename memory files

Usage

import { ChatAnthropic, tools } from "@langchain/anthropic";

const llm = new ChatAnthropic({
  model: "claude-sonnet-4-5-20250929",
  clientOptions: {
    defaultHeaders: {
      "anthropic-beta": "context-management-2025-06-27",
    },
  },
});

const memory = tools.memory_20250818({
  execute: async (action) => {
    // Handle memory command execution with your storage backend
    // ...
  },
});

const llmWithMemory = llm.bindTools([memory]);
const response = await llmWithMemory.invoke("Remember that I like Python");

@pokey has been attempting to implement the Anthropic memory tool as part of a middleware. This PR approaches this differently and implements this tool as provider tool. Running the tool within and agent shows that the model can interact with the tool:

{
  messages: [
    HumanMessage {
      "id": "b107d300-b413-41b4-9047-06b5e41fc063",
      "content": "What is the weather in Tokyo?",
      "additional_kwargs": {},
      "response_metadata": {}
    },
    AIMessage {
      "id": "msg_01CEiqjq2L7vATupz85S17CS",
      "content": [
        {
          "type": "text",
          "text": "I'll check my memory first, then help you with the weather information."
        },
        {
          "type": "tool_use",
          "id": "toolu_01W4GpFJk6wiEhPvrZy5efLP",
          "name": "memory",
          "input": {
            "command": "view",
            "path": "/memories"
          }
        }
      ],
      "name": "model",
      "additional_kwargs": {...},
        "context_management": {
          "applied_edits": []
        }
      },
      "response_metadata": {...},
      "tool_calls": [
        {
          "name": "memory",
          "args": {
            "command": "view",
            "path": "/memories"
          },
          "id": "toolu_01W4GpFJk6wiEhPvrZy5efLP",
          "type": "tool_call"
        }
      ],
      "invalid_tool_calls": [],
      "usage_metadata": {...}
    },
    ToolMessage {
      "id": "c4ce0fb9-824c-4ca8-ae35-796e201d25c1",
      "content": "Directory: /memories\n- weather_in_tokyo.txt",
      "name": "memory",
      "additional_kwargs": {},
      "response_metadata": {},
      "tool_call_id": "toolu_01W4GpFJk6wiEhPvrZy5efLP"
    },
    AIMessage {
      "id": "msg_015tZU7NaGgTRvKTUEUAq8Ls",
      "content": [
        {
          "type": "text",
          "text": "Let me check what information I have about Tokyo's weather:"
        },
        {
          "type": "tool_use",
          "id": "toolu_017gEW3kSusW1SotXxEdbEFX",
          "name": "memory",
          "input": {
            "command": "view",
            "path": "/memories/weather_in_tokyo.txt"
          }
        }
      ],
      "name": "model",
      "additional_kwargs": {...},
      "response_metadata": {...},
      "tool_calls": [
        {
          "name": "memory",
          "args": {
            "command": "view",
            "path": "/memories/weather_in_tokyo.txt"
          },
          "id": "toolu_017gEW3kSusW1SotXxEdbEFX",
          "type": "tool_call"
        }
      ],
      "invalid_tool_calls": [],
      "usage_metadata": {
        "input_tokens": 1689,
        "output_tokens": 90,
        "total_tokens": 1779,
        "input_token_details": {
          "cache_creation": 0,
          "cache_read": 0
        }
      }
    },
    ToolMessage {
      "id": "a477ec43-f8e8-4e8a-b5d0-9b2ab25e0384",
      "content": "1|{\"weather\":\"sunny\",\"temperature\":20,\"humidity\":50,\"wind_speed\":10,\"wind_direction\":\"N\",\"wind_gust\":15,\"wind_gust_direction\":\"N\",\"wind_gust_speed\":20,\"time\":\"2025-11-14T07:41:48.543Z\"}",
      "name": "memory",
      "additional_kwargs": {},
      "response_metadata": {},
      "tool_call_id": "toolu_017gEW3kSusW1SotXxEdbEFX"
    },
    AIMessage {
      "id": "msg_01UNpzE3YXH44wfWkNTkRtfH",
      "content": "I apologize, but I don't have access to real-time weather data or weather APIs. The information in my memory appears to be from a previous interaction and is not current.\n\nTo get accurate, up-to-date weather information for Tokyo, I recommend:\n\n1. **Weather websites**: Check weather.com, accuweather.com, or weather.gov\n2. **Search engines**: Simply search \"Tokyo weather\" on Google or Bing\n3. **Weather apps**: Use apps like Weather Channel, AccuWeather, or your phone's built-in weather app\n4. **Japan Meteorological Agency**: Visit their official website at jma.go.jp for official forecasts\n\nThese sources will provide you with current conditions, temperature, humidity, precipitation, and forecasts for Tokyo.",
      "name": "model",
      "additional_kwargs": {...},
      "response_metadata": {...},
      "tool_calls": [],
      "invalid_tool_calls": [],
      "usage_metadata": {...}
    }
  ]
}

[changeset-bot]

⚠️ No Changeset found

Latest commit: 03204b0

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

This PR includes no changesets

When changesets are added to this PR, you'll see the packages that this PR includes changesets for and the associated semver types

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[pokey]

Nice! I guess what I'm not following here is how the files end up on the graph state. In your example, it looks like the files just end up in an object that's a global variable declared in the example file. Is there some magic I'm missing here to propagate this to / from the graph state?

[christian-bromann]

Is there some magic I'm missing here to propagate this to / from the graph state?

If your goal is it to store the file system in the graph as part of your agent you could do this via:

import { z } from "zod";
import { ChatAnthropic, tools, StateFileSystem } from "@langchain/anthropic";

import { createAgent, createMiddleware } from "langchain";

const model = new ChatAnthropic({
  model: "claude-sonnet-4-5-20250929",
  temperature: 0,
  clientOptions: {
    defaultHeaders: {
      "anthropic-beta": "context-management-2025-06-27",
    },
  },
});

const memoryMiddleware = () => {
  const files = {};
  const filesystem = new StateFileSystem(files, [], (files) => {
    Object.assign(files, files);
  });

  return createMiddleware({
    name: "MemoryMiddleware",
    stateSchema: z.object({
      files: z.record(
        z.string(),
        z.array(
          z.object({
            content: z.string(),
            createdAt: z.string(),
          })
        )
      ),
    }),
    tools: [tools.memory({ filesystem })],
    beforeAgent: async (state) => {
      // update the filesystem with files within the state
      await Promise.all(
        Object.entries(state.files).map(async ([filepath, fileData]) =>
          filesystem.writeFile(filepath, fileData)
        )
      );
    },
  });
};

const agent = createAgent({
  model,
  middleware: [memoryMiddleware],
});

const result = await agent.invoke({
  messages: "What is the weather in Tokyo?",
});

console.log(result);

[pokey]

Hmm still not sure I'm following. I see how the filesystem gets what's initially on the graph, but I'm still not seeing how it updates the graph state after the tool call

[christian-bromann]

It's part of the tool implementation in the provider package. By passing in a FileSystem type (either memory of fs style filesystem) it will be attached to the tool call and handled as part of the tool implementation.

[pokey]

Maybe worth an integration test or two to show how that end-to-end pipeline works? Could just use / adapt mine

[christian-bromann]

Decided to move the memory command handler out of this PR and will make it part of a middleware