- In short
- Every Claude conversation has a finite working-memory budget -- the context window -- that fills as messages and uploaded documents accumulate. As the window nears its limit, the platform can automatically summarise earlier messages to make room, under specific plan and feature conditions. Summarisation compresses detail rather than deleting it outright, and full history typically remains available for reference, so a long session can behave differently late on than it did early purely because of this budget effect.
A budget you spend by talking
Every Claude conversation runs against a finite working-memory budget called the context window, and the CCAO-F exam introduces it at the understand level because so much of context management follows from grasping this one fact. The window is not unlimited scratch space; it is a fixed budget that fills as messages and uploaded documents accumulate. Every message you send and every document you attach spends a little of it.
The consequence that matters is that a long session is not a neutral, unchanging container. As the budget fills, the platform may take action to keep the conversation going, and that action changes how the session behaves. Understanding the window as a finite budget is the foundation for recognising context degradation and for the restart, summarise, or persist responses that follow.
- The context window as a finite budget
- The finite working-memory budget of a Claude conversation, which fills as messages and uploaded documents accumulate. As it nears its limit, the platform can automatically summarise earlier messages to make room, under specific plan and feature conditions. Summarisation compresses detail rather than deleting it, and full history typically remains available, so a long session can behave differently late on than early.
Filling and summarising
The window fills steadily as the exchange grows. A short conversation uses little of the budget; a long one, especially with large uploaded documents, consumes much more. When the window nears its limit, the platform can automatically summarise earlier messages to make room so the session can continue rather than simply stopping. This automatic summarisation is what keeps a long conversation usable past the point where raw accumulation would otherwise fill the budget.
The important nuance is what summarisation does and does not do. It compresses detail rather than deleting content outright, and the full history typically remains available for reference. So earlier material is not erased; it is condensed. That distinction matters because compression, not deletion, is what produces the characteristic late-session behaviour changes -- an instruction is still "there" in history, but its detail has been compressed in the working context, which can weaken how strongly it is followed.
Conditional, not guaranteed
A point the exam is careful about is that automatic summarisation is tied to specific plan and feature conditions rather than being a universal guarantee. It is not the case that every account, on every plan, with every feature configuration, gets automatic summarisation in the same way. Whether and how the platform summarises as the window fills depends on the conditions in effect for that account.
This conditionality is why two mistaken assumptions are both worth avoiding. One is treating context capacity as effectively unlimited -- it is finite on every plan. The other is assuming automatic summarisation is guaranteed everywhere -- it is conditional. The safe mental model holds both: the window is always finite, and the platform's automatic help with a full window is something that applies under particular conditions, not a fixed feature to rely on unconditionally. This mirrors how other platform behaviours are plan-dependent and worth verifying rather than assuming.
What the CCAO-F exam trips candidates on
Two assumptions are tested. The first is assuming context capacity is effectively unlimited on any plan or configuration. Every conversation's window is finite, and a long enough session will press against it regardless of plan. The credited understanding treats the budget as real and bounded.
The second is assuming automatic summarisation is guaranteed in every account and feature configuration rather than tied to specific conditions. A question may imply the platform will always seamlessly summarise a full window; the credited reading notes that this behaviour is conditional on plan and features, so it should not be assumed universally.
Worked example
A user runs a two-hour session, pasting long documents throughout, and is surprised that Claude's later responses seem to follow the early instructions less closely. A colleague reassures them: 'The context window is basically unlimited, and Claude always summarises automatically so nothing is ever lost.' What in that reassurance is wrong?
Both halves of the reassurance rest on the mistaken assumptions the exam targets. The window is not "basically unlimited" -- it is a finite budget, and a two-hour session with long documents pasted throughout is exactly the kind of exchange that fills it. The very symptom the user noticed, early instructions being followed less closely later, is the fingerprint of a budget that has filled and been compressed, not of an unlimited container.
The claim that Claude "always summarises automatically so nothing is ever lost" is wrong in two ways. First, automatic summarisation is conditional on plan and features, not a universal guarantee, so it cannot be assumed to be in effect for every account. Second, even where it operates, summarisation compresses detail rather than preserving everything at full fidelity -- that is precisely why the early instructions lost some force. Full history typically remains available for reference, so the content is not deleted, but the working detail is condensed, which is what changed the late-session behaviour.
The accurate picture is that the finite budget filled over two hours, compression kicked in (if the conditions applied), and the compression is why the later responses drifted from the early instructions. Nothing is malfunctioning; the budget effect is doing exactly what it does. The user should treat the window as finite and plan for it, not assume it is limitless and self-healing.
Common misreadings to avoid
Misconception
The context window is effectively unlimited, so long sessions behave the same throughout.
What's actually true
Misconception
Claude always summarises a full window automatically, so nothing is ever lost.
What's actually true
How this shows up on the exam
Questions describe a long session behaving differently late on and ask what is happening, or bait with claims that context is unlimited or summarisation is guaranteed. Anchor on the finite budget: the window fills, the platform may summarise under specific conditions, and compression -- not deletion -- explains the change. The reliable answer treats capacity as bounded and summarisation as conditional.
This foundational knowledge point unlocks recognising context degradation and underpins the three responses and the long-session diagnosis.
During a long session with many pasted documents, Claude's later replies follow the early instructions less closely. Which explanation is most accurate?
People also ask
What is the context window in Claude?
What happens when the context window fills up?
Is context capacity unlimited on any plan?
Watch and learn
Official Anthropic Academy lessons first, then hand-picked walkthroughs. Videos load only when you press play.
No videos curated for this concept yet
We are still curating the best official and community videos for this topic.
Official prep for this domain
Anthropic's own free prep module for this part of the syllabus, on the official prep course. Free with an Anthropic Academy sign-in.
References & primary sources
Master this concept with Archie
Practice it inside an adaptive study session. Archie, your Socratic AI tutor, tracks your mastery with Bayesian Knowledge Tracing and schedules the perfect next review.