OpenClaw Context Explained

Context is one of those OpenClaw words that sounds fuzzy until you read the docs carefully. OpenClaw uses it in a very literal way. Context is everything the runtime sends to the model for a run, and it is limited by the model's context window. That means your chat history is only one part of the story. The system prompt, tool schemas, workspace bootstrap files, tool results, attachments, and compaction summaries all compete for the same budget.

What actually sits inside context

The docs break the current run into a few big pieces. There is the OpenClaw-built system prompt, which includes rules, available tools, skills metadata, runtime details, and injected workspace files. Then there is the conversation history for the session, followed by tool calls and tool results. If you read a big file, attach an image, or run a command with a noisy output, that also becomes part of the budget. In other words, context is not just words you typed. It is the whole working set the model receives.

That is also why context is not the same thing as memory. Memory can live on disk in files and come back later. Context is whatever has been loaded into the current turn. The docs are pretty direct about this boundary because operators often expect OpenClaw to have a magical long-term memory store. It does not. It has a context window, plus files and session history that can be reintroduced into that window.

The commands that make context visible

OpenClaw ships with a solid inspection loop. /status gives you a quick sense of how full the window is. /context list shows what was injected and rough sizes by file and total. /context detail goes deeper and breaks down file sizes, skill entry sizes, the system prompt, and tool schema overhead. /usage tokens can append usage data to replies. When the conversation starts getting too heavy, /compact summarizes older history so the chat can keep moving.

/status
/context list
/context detail
/usage tokens
/compact

Why bootstrap files matter so much

The docs also explain that OpenClaw injects a fixed set of workspace files by default when they exist, including AGENTS.md, SOUL.md, TOOLS.md, IDENTITY.md, USER.md, HEARTBEAT.md, and BOOTSTRAP.md on first run. These are not free. Large files are truncated per file with bootstrapMaxChars and capped again by bootstrapTotalMaxChars. That is a subtle but important operator lesson. If you stuff huge amounts of loose notes into boot files, you are spending context before the model has even read the latest message.

Skills have a similar split. The skills list itself is injected as compact metadata, so every available skill carries some overhead. The full SKILL.md instructions are not injected by default. The model is expected to read the chosen skill on demand. Tools create two kinds of overhead too: the visible tool list text and the tool JSON schemas sent to the model. If you are trying to understand why a session feels crowded, those hidden costs matter.

The practical operator takeaway

Once you see context as a budget instead of a vibe, OpenClaw makes more sense. Keep bootstrap files deliberate, keep noisy tool output under control, use /context when a session feels bloated, and compact older history before it becomes a crisis. The docs are basically teaching an economics lesson: every extra token has an opportunity cost. Operators who treat context as a finite working memory end up with more stable runs, cleaner prompts, and fewer surprises when conversations get long.

If you want the operator version of these docs turned into a practical working system, read The OpenClaw Playbook. It connects official OpenClaw features to real workflows, guardrails, and deployment decisions.