Context Assembly: Files, History, Instructions, and Tool Schemas

Many agent problems are not caused by a weak model. They come from bad context assembly.

The model acts on what it currently sees.

The OpenClaw context docs define context clearly:

Context is everything OpenClaw sends to the model for a run.

The Key Idea: Context Is Not Memory

Context includes:

System prompt
Conversation history
Tool calls and results
Attachments
Compaction summaries
Injected workspace files
Tool schemas

Memory may live on disk and be retrieved later.

Context is what enters the model window now.

System Prompt Is Rebuilt Each Run

OpenClaw owns and rebuilds the system prompt for each run. It usually includes:

tool list
skills list
workspace location
time and runtime metadata
Project Context injected files

It is not just text appended by the user. It is assembled from several runtime sources.

Project Context

OpenClaw injects workspace bootstrap files such as:

AGENTS.md
SOUL.md
TOOLS.md
IDENTITY.md
USER.md
HEARTBEAT.md
BOOTSTRAP.md

Large files are truncated. /context list shows raw size, injected size, and truncation status.

Write these files carefully:

clear rules
accurate paths
no huge logs
no full chat history as permanent context

Tools Have Two Context Costs

Tools cost context in two ways:

tool list text
  appears in system prompt so the model knows capabilities

tool JSON schemas
  sent to the model so it can call tools correctly

Many people inspect chat history but forget tool schemas. A large tool catalog can consume context before the model starts answering.

History, Compaction, Pruning

Conversation history enters context, but it cannot grow forever.

OpenClaw uses:

compaction
  summarize old history into a compact entry

pruning
  remove old tool results from current prompt without rewriting disk transcript

So "it used to know this" may mean the information did not enter current context.

Inspecting Context

Useful commands:

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

Check:

system prompt size
injected workspace files
truncated files
tool schema size
history size
window pressure

Common Misunderstandings

Misunderstanding 1: Memory Equals Context

No. Memory is stored or retrievable; context is the current model window.

Misunderstanding 2: More Files Is Better

Not always. Irrelevant information dilutes the prompt.

Misunderstanding 3: More Tools Always Means More Power

More tools also mean more schema cost and higher accidental-use risk.

Final Summary

Context engineering decides what the model sees.

In one sentence:

Agent behavior = model capability + current context + available tools + permission boundaries.

Lesson Homework

1. Run /context detail in one session.
2. Identify the three biggest context contributors.
3. Check whether TOOLS.md is too long.
4. Run /compact and observe window pressure.

Next Lesson Preview

Next: tool call protocol and how models decide which tool to call.

References

• OpenClaw Docs: Context
• OpenClaw Docs: Context engine
• OpenClaw Docs: System prompt
• OpenClaw Docs: Compaction
• OpenClaw Docs: Session pruning

Original link: Context Assembly: Files, History, Instructions, and Tool Schemas