CORE GUIDE
Context window
The context window is a finite working-input budget shared by instructions, conversation, retrieved evidence and tool results.
Mental model
Context is the model's current workspace, not an infinite memory. Every additional token competes for limited attention and capacity.
Why it matters
A larger context window does not remove the need for context engineering. Long histories can bury constraints, stale information can conflict with current evidence, and tool traces can crowd out the information needed for the next decision.
01
What competes for the window
System instructions, user messages, examples, retrieved documents, memory snippets and tool outputs all occupy the same finite input. As the workspace grows, the application must decide what to preserve, compress, retrieve again or discard.
02
Example: an agent after 80 tool calls
An agent keeps every tool result in the conversation. By the final step, the original acceptance criteria and recent authoritative state are surrounded by thousands of low-value log tokens. The fix is not simply a bigger model; it is a policy that retains decisions and evidence while compacting transient traces.
Common failure modes
- Treating more context as automatically better context.
- Keeping stale or contradictory evidence indefinitely.
- Using chat history as a substitute for durable workflow state.
Engineering heuristics
- Allocate the context budget by role and importance.
- Prefer current authoritative evidence over old verbose traces.
- Compact with explicit retention rules for constraints, decisions and unresolved questions.
Takeaways
- 01Context is a finite workspace.
- 02Relevance and authority matter more than raw volume.
- 03Durable state and context are different system responsibilities.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.