If your LLM stops responding, it’s not moody—it’s out of context window. Use /compact to summarize the conversation, reclaim tokens, and continue without losing the thread.
Why Is This Happening?
Every LLM has a Context Window—its short‑term memory. Each message and response consumes tokens. When the total exceeds the model’s limit, the AI can’t “read” the entire thread, and responses stall or error.
The Solution: The /compact Command
OpenClaw includes a built‑in feature to fix this without starting a brand‑new chat. /compact summarizes the conversation into a concise state and replaces the oversized history with that summary.
How to use it
- Identify the Stall: If you see “Context Limit” or mid‑sentence stops, focus the input box.
- Type the Command: Enter
/compactand press Enter. - The Magic: OpenClaw asks the LLM to generate a concise summary of key points, decisions, and context.
- Memory Reset: The large message history is replaced with this single summary.
- Result: You keep the “soul” of the session and regain 80–90% of the context window.
/compactPro‑Tips for Managing Context
Be Proactive
Don’t wait for errors. If the thread is getting long, run /compact before the model slows down.
Selective Pruning
Manually delete redundant code blocks or “thank you” messages. Every token counts in long sessions.
Switch Models
When possible, use a long‑context model (e.g., Claude 3.5 Sonnet, GPT‑4o) if your API credits allow.
Troubleshooting: What if /compact Fails?
If the context is so full the model can’t even process a summary request:
- Manual Deletion: Delete the last 5–10 long messages to free enough room for /compact to run.
- Fresh Start: Copy the most important requirements into a new chat thread and continue there.