Conversation Context and Memory Management Strategies

The Challenge of Context Management

AI models understand conversation history through the context window. A longer context yields more coherent answers, but also increases token consumption. OpenClaw provides multiple strategies to balance quality and cost.

Basic Configuration

{
  "sessions": {
    "maxHistory": 20,
    "maxTokens": 50000,
    "contextStrategy": "sliding-window"
  }
}

Context Strategies

Sliding Window

Keeps the most recent N turns of conversation, discarding older history:

{
  "sessions": {
    "contextStrategy": "sliding-window",
    "maxHistory": 20
  }
}

Pros: Simple, predictable memory usage Cons: May lose important early context

Smart Trim

Retains context based on relevance:

{
  "sessions": {
    "contextStrategy": "smart-trim",
    "maxTokens": 50000,
    "relevanceThreshold": 0.5
  }
}

How it works: Computes the relevance of each historical message to the current conversation, keeps highly relevant messages, and trims irrelevant ones from the middle.

Summary Mode

Automatically generates a summary when the conversation exceeds a threshold:

{
  "sessions": {
    "contextStrategy": "summary",
    "summaryAfter": 15,
    "summaryModel": "fast",
    "summaryPrompt": "Summarize the following conversation into concise bullet points, preserving key information and user preferences.",
    "keepRecent": 5
  }
}

How it works:

1. Summary is triggered at 15 turns
2. A lightweight model generates the conversation summary
3. The summary replaces the old conversation history
4. The 5 most recent turns are kept as-is

Hybrid Mode

Combines multiple strategies:

{
  "sessions": {
    "contextStrategy": "hybrid",
    "phases": [
      {"maxMessages": 10, "strategy": "full"},
      {"maxMessages": 30, "strategy": "smart-trim"},
      {"maxMessages": 999, "strategy": "summary"}
    ]
  }
}

Full history for the first 10 turns, smart trimming for turns 10-30, and summarization beyond 30.

System Prompt Optimization

System prompts also consume context space and should be kept concise:

{
  "agents": {
    "main": {
      "systemPrompt": "You are an AI assistant. Be concise; elaborate only when necessary.",
      "dynamicPrompt": {
        "base": "You are an AI assistant.",
        "additions": [
          {"condition": "tools_available", "text": "You can use the following tools: ..."},
          {"condition": "user_is_new", "text": "This is a new user. Please guide them kindly."}
        ]
      }
    }
  }
}

Dynamic prompts only add extra content when needed, reducing unnecessary token consumption.

Combining with Vector Memory

Store long-term information in vector memory, freeing it from the context:

{
  "sessions": {
    "autoMemorize": {
      "enabled": true,
      "triggerWords": ["remember", "from now on", "always"],
      "extractFacts": true
    },
    "autoRecall": {
      "enabled": true,
      "topK": 3,
      "threshold": 0.8
    }
  }
}

When a user says "Remember that I prefer concise answers," this preference is stored in vector memory and no longer needs to be kept in the conversation history.

Context Budget

Allocate token budgets for different components:

{
  "sessions": {
    "tokenBudget": {
      "total": 100000,
      "systemPrompt": 2000,
      "memory": 3000,
      "history": 90000,
      "tools": 5000
    }
  }
}

Monitor Context Usage

openclaw sessions context-stats

Context Usage Statistics:
  Avg context length: 12,500 tokens
  Max context length: 85,000 tokens
  Avg history messages: 15
  Context overflow events: 3 (auto-trimmed)
  Estimated context cost: $1.25/day

Per-Channel Customization

Different channels can use different context strategies:

{
  "channels": {
    "telegram-main": {
      "session": {
        "maxHistory": 30,
        "contextStrategy": "hybrid"
      }
    },
    "whatsapp-quick": {
      "session": {
        "maxHistory": 5,
        "contextStrategy": "sliding-window"
      }
    }
  }
}

Customer support scenarios benefit from longer history to understand the full issue, while quick Q&A scenarios need only minimal context.

Manual Management

Users can manage their own conversation context through commands:

User: /clear
AI: Conversation history cleared.

User: /context
AI: The current conversation contains 15 messages, using 8,500 tokens.

{
  "channels": {
    "telegram-main": {
      "commands": {
        "/clear": "session.clear",
        "/context": "session.info"
      }
    }
  }
}

Summary

Context management is a critical part of OpenClaw performance and cost optimization. Choose the right strategy for each scenario -- sliding window for simple chats, summary mode for long-running services, vector memory for unlimited long-term recall -- to maintain a great conversational experience while effectively controlling token costs.