The Context Window

Models have a limited memory span (Context Window).

What fills it up?

1. System Prompt: Instructions (2k tokens)
2. History: Recent chats (5k+ tokens)
3. Files: Read files (context.read) (Infinite potential)

Compaction

When context gets full (e.g., >100k tokens), OpenClaw triggers Compaction. It summarizes the oldest messages into MEMORY.md and deletes the raw logs to free up space.