Claude context management: what breaks in a long session, and what the exams ask

Claude context management is the discipline of keeping a long session accurate about what it already established. Histories get compressed, and what goes is consistent: the topic stays and the specifics leave — the figure, the date, the exact commitment. Nothing warns you when it happens, so it surfaces later as a confident, specific answer that is wrong. The examinable fix is almost never a better prompt; it is holding what must stay exact outside the part that compresses.

Where claude context management sits on the blueprint

It is a domain in its own right on Architect Foundations — Context Management and Reliability, 15%, about nine of sixty questions. On Architect Professional the neighbouring domain, Claude Models, Prompting and Context Engineering, is 13%.

Those figures understate it. Because the material is about what happens to long-running work, it also shows up inside scenarios belonging to other domains: an orchestration question where a subagent returns too much, an error-handling question where a coordinator has lost track of what succeeded. The published domain weights put a floor under how much it is worth, not a ceiling.

Compression drops particulars, not topics

The loss follows a pattern, and knowing the pattern is most of the skill. What a conversation was about comes through fine. The specifics inside it do not: an amount becomes “the refund”, a date becomes “recently”, and a commitment made in precise terms becomes a general sense that something was promised.

That table is also the answer key for a whole family of exam items. When a scenario describes a long-running assistant getting a figure subtly wrong, or reviving an approach that was explicitly rejected earlier, the correct option moves that information out of the conversation. The distractors rewrite the summarisation instruction, or ask the model to be more careful.

Position inside a long input is a design decision

A second effect compounds the first. Whatever sits at the two ends of a long input gets read more dependably than whatever is stacked in between, so placement changes how likely something is to be used. That is not a quirk to work around; it is a layout decision you own.

The same reasoning governs tool output. A lookup that returns forty fields when the task uses five does not simply waste budget once — it sits there diluting everything around it on every later turn. Cutting it down on arrival costs less and holds better than squeezing it afterwards, because the cost is never incurred in the first place.

Long sessions drift toward the typical

This is the failure candidates find hardest to name, and it is worth learning as a phrase. After a long session, answers start coming from how things of this kind usually look, rather than from the particular one the agent examined an hour earlier. It will describe a typical arrangement with the same confidence it used for something it genuinely read.

Three fixes, and they compose

Write findings down as you establish them. A record consulted later beats recall, because it survives whatever the conversation does to itself. This is the cheapest of the three and the one most often skipped.

Push bulky work somewhere else. Searching a codebase generates volumes of raw material for a handful of conclusions, so running it in a separate agent leaves the coordinating context holding structure instead. That is an orchestration decision in its own right — see what a subagent must be told and when the extra structure is worth it.

Close each phase before opening the next. Summarise deliberately at a natural boundary, while you can still choose what carries forward. Compression forced at the limit happens mid-step, on terms you did not set, and cannot know which earlier findings the rest of the job still needs.

Resuming is not automatically the right move

A resumed session brings back its accumulated state — results it collected, files it read, conclusions it reached — and reasons over all of it with the authority of something that has genuinely done the work. That is the benefit and the hazard in one sentence.

Where a lot has moved underneath it, or where the session settled on something that turned out to be false, carrying that snapshot forward is weaker than starting again from a written statement of the current position. A correction does not delete the original; it adds a note beside something the session still holds as established. The trade is real in both directions: a fresh session contains exactly what you chose to write down, and genuinely loses whatever you left out.

For the configuration side of long-running work, see Claude Code configuration, and if you have not picked a track, the certification comparison guide covers the differences. Official outlines are published on the Anthropic Partner Academy certification pages.

Key takeaways

• Compression is lossy in a known direction. Subjects survive; amounts, dates, identifiers and exact commitments do not.
• Move what must stay exact out of the conversation. A structured record travelling with each request cannot be reached by compressing the history.
• Placement is a design choice. Openings and endings are attended to reliably, so lead with what matters and structure the rest.
• Late in a session, answers slide toward the generic — silently. Prefer designs that hold without anyone noticing, because nobody will.
• A bigger window postpones the problem. On these exams, enlarging the window in answer to a structural fault is usually the distractor.
• Resuming carries stale state as confidently as fresh state. When the world has moved, a session started clean from a written summary is often the safer option.