Function contextEditingMiddleware

contextEditingMiddleware(
config?: ContextEditingMiddlewareConfig,
): AgentMiddleware<undefined, undefined, any>

Middleware that automatically prunes tool results to manage context size.

This middleware applies a sequence of edits when the total input token count exceeds configured thresholds. By default, it uses the ClearToolUsesEdit strategy which mirrors Anthropic's clear_tool_uses_20250919 behaviour by clearing older tool results once the conversation exceeds 100,000 tokens.

Basic Usage

Use the middleware with default settings to automatically manage context:

Parameters

config: ContextEditingMiddlewareConfig = {}
Configuration options for the middleware

Returns AgentMiddleware<undefined, undefined, any>

A middleware instance that can be used with createAgent

Example: Basic usage with defaults

import { contextEditingMiddleware } from "langchain";
import { createAgent } from "langchain";

const agent = createAgent({
  model: "anthropic:claude-sonnet-4-5",
  tools: [searchTool, calculatorTool],
  middleware: [
    contextEditingMiddleware(),
  ],
});

The default configuration:

Triggers when context exceeds 100,000 tokens
Keeps the 3 most recent tool results
Uses approximate token counting (fast)
Does not clear tool call arguments

Custom Configuration

Customize the clearing behavior with ClearToolUsesEdit:

Example: Custom ClearToolUsesEdit configuration

import { contextEditingMiddleware, ClearToolUsesEdit } from "langchain";

// Single condition: trigger if tokens >= 50000 AND messages >= 20
const agent1 = createAgent({
  model: "anthropic:claude-sonnet-4-5",
  tools: [searchTool, calculatorTool],
  middleware: [
    contextEditingMiddleware({
      edits: [
        new ClearToolUsesEdit({
          trigger: { tokens: 50000, messages: 20 },
          keep: { messages: 5 },
          excludeTools: ["search"],
          clearToolInputs: true,
        }),
      ],
      tokenCountMethod: "approx",
    }),
  ],
});

// Multiple conditions: trigger if (tokens >= 50000 AND messages >= 20) OR (tokens >= 30000 AND messages >= 50)
const agent2 = createAgent({
  model: "anthropic:claude-sonnet-4-5",
  tools: [searchTool, calculatorTool],
  middleware: [
    contextEditingMiddleware({
      edits: [
        new ClearToolUsesEdit({
          trigger: [
            { tokens: 50000, messages: 20 },
            { tokens: 30000, messages: 50 },
          ],
          keep: { messages: 5 },
        }),
      ],
    }),
  ],
});

// Fractional trigger with model profile
const agent3 = createAgent({
  model: chatModel,
  tools: [searchTool, calculatorTool],
  middleware: [
    contextEditingMiddleware({
      edits: [
        new ClearToolUsesEdit({
          trigger: { fraction: 0.8 },  // Trigger at 80% of model's max tokens
          keep: { fraction: 0.3 },     // Keep 30% of model's max tokens
          model: chatModel,
        }),
      ],
    }),
  ],
});

Custom Editing Strategies

Implement your own context editing strategy by creating a class that implements the ContextEdit interface:

Example: Custom editing strategy

import { contextEditingMiddleware, type ContextEdit, type TokenCounter } from "langchain";
import type { BaseMessage } from "@langchain/core/messages";

class CustomEdit implements ContextEdit {
  async apply(params: {
    tokens: number;
    messages: BaseMessage[];
    countTokens: TokenCounter;
  }): Promise<number> {
    // Implement your custom editing logic here
    // and apply it to the messages array, then
    // return the new token count after edits
    return countTokens(messages);
  }
}