Context

Context is the complete set of information sent to the LLM for a single inference turn. It is the “world” the AI sees at that moment.

A typical context is built from layers, ordered from general to specific:

• Role: Defines who the agent is.
• Content: “You are OpenClaw, a helpful AI assistant. You answer concisely…”
• Priority: High. This sets the behavioral baseline.

• Time: “Current date: 2023-10-27”.
• User Info: “User name: Alice. Location: New York.”
• OS/Platform: “Running on Linux. Channel: Telegram.”

• Schema: JSON descriptions of available functions (e.g., web_search, read_file).
• Instruction: “Use web_search if the user asks for current events.”

• Retrieval: Relevant snippets fetched from the vector database based on the user’s latest query.
• Content: “Alice mentioned she likes sushi in a previous chat.”

• The Chat: The actual back-and-forth messages.
• [User]: Hi!
• [Agent]: Hello! How can I help?
• [User]: What's the weather?
• Compaction: Older history may be replaced by a summary here.

• Chain-of-Thought: If the agent is “thinking”, previous reasoning steps are included here to guide the final answer.

The Context Window is the maximum size of this combined text (measured in tokens).

• GPT-4: 8k or 32k or 128k.
• Claude 3: 200k.
• Llama 3: 8k/128k.

OpenClaw dynamically manages this budget:

1. Reserve space for the System Prompt and Tools (must always fit).
2. Fill with History (most recent first).
3. Inject RAG memories if space permits.
4. Truncate or Compact the oldest history if the limit is reached.

To see exactly what the agent sees, you can often run with verbose logging:

openclaw run --verbose

This will print the full prompt sent to the LLM, allowing you to inspect if memories are being retrieved or if the history is being truncated correctly.