# Context window management

Actively curating what goes into the model's finite token window so the relevant signal isn't crowded out.

## Why the CCAR-P exam tests this

The exam checks you treat the context window as a finite, curated budget rather than dumping everything in.

## Definition

The **context window** is the finite number of tokens the model can attend to in one request — system prompt, conversation history, retrieved documents, tool results, and the model's own output all share it. Everything you include competes for that budget and adds cost and latency.

Good management means being **selective**: retrieve and include only the passages relevant to the task, summarize or truncate old history, and drop stale tool output. "Stuff the whole knowledge base into the prompt" is an anti-pattern — it raises cost, slows the call, and can bury the signal the model needs.

Management techniques include **RAG** (fetch just the relevant chunks), **summarization/compaction** of long histories, and ordering content so the most important material is clearly placed. Larger windows raise the ceiling but do not remove the need to curate — more tokens still cost more and can dilute attention.

## Exam trap

A bigger context window is not a substitute for curation; overloading it with irrelevant tokens raises cost and latency and can degrade answer quality (see context rot).

## Commonly confused with

- **context-rot** — Context rot is the quality decay that happens when the window is poorly managed and stuffed with low-value tokens.
- **prompt-caching** — Caching cheapens reprocessing of a prefix but does not shrink how much of the window it occupies.

## Related

- context-rot
- prompt-caching
- no-customer-fine-tuning

---

Source: Claude Cert Prep — an independent, unofficial CCAR-P study resource (domain P2). Not affiliated with Anthropic.