Rubric
Contents — domains, guide and mocks

Context limits and memory

CCAO-F 3.415 min read · checked 21 September 2026

Task statementUnderstand and manage context limitations and memory considerations (when to restart, summarize, or persist)

What is on the table

all of it inside one fixed window

  1. Project instructionsthe standing brief, on every turn
  2. Retrieved knowledgethe relevant parts of the knowledge base
  3. Attachmentsthe files you uploaded in this conversation
  4. Conversation so farevery message, yours and Claude’s
  5. Your newest messagethe one you just sent
  6. Reserved for the answerspace held back for the response
Everything here competes for the same fixed space, and the reserve at the bottom is why the usable conversation is slightly smaller than the headline number.

What the context window is, and how big

The context window is the total amount of text the model can hold in view at once — your messages, its replies, the files you attached, and whatever the Project supplied. It is not a per-message limit and it is not storage. When it is full, something has to give.

Size depends on the model and the plan. Anthropic’s help centre states that on paid plans newer models support up to one million tokens, while others support five hundred thousand or two hundred thousand, and that part of the window is reserved for Claude’s response, making the maximum conversation slightly smaller than the full window. The models documentation records the same ordering on the technical side: the current larger models document a one-million-token window, while the fastest tier documents two hundred thousand. Tokens are roughly word-sized chunks, so a very long report can occupy a serious fraction of a window on its own.

What happens as a conversation gets long

You do not usually hit a wall. As a conversation approaches the context limit, Claude automatically summarises earlier messages so it can keep going, and the help centre notes you may see Claude “organising its thoughts” while this happens. Your full chat history is preserved, so earlier content can still be referred to afterwards. This automatic context management depends on code execution being enabled, and conversations that rely on it consume more of your usage limit. Rare edge cases — a very large first message, for example — can still run into the limit.

Two consequences matter for everyday work. First, an instruction you gave an hour and sixty messages ago may now be part of a summary rather than sitting there verbatim, which is why long threads quietly stop obeying a constraint set at the start. If a rule matters, restate it at the point where it bites. Second, a long conversation is an expensive one: every turn carries the accumulated history, and the help centre’s own advice is to try starting a new conversation if you are approaching your usage limit in a longer chat.

Three moves, and when each is right

Restart, summarise and persist are not alternatives to one another so much as a sequence. Summarising is how you leave a conversation well; restarting is where you go; persisting is how you avoid doing it again next week.

Which move does this situation call for?

Why is the conversation struggling?
  • Long, tangled, drifting
    Summarise, then restartcarry forward the conclusions
  • New topic, same day
    Just start a new chatnothing worth carrying
  • Same context every week
    Persist it in a Projectknowledge base and instructions
  • A stable fact about you
    Let memory hold itrole, preferences, ongoing work
The middle branch is the one people skip — they restart without summarising, and throw away the decisions that took an hour to reach.

A good summary before a restart is not “summarise this conversation”. It is a handoff note: the decisions made and still standing, the constraints that must survive, the current best version of the work, and the open questions. Ask for exactly those four things, read it, correct it — a summary is an output like any other and can drop the one decision you cared about — and paste it as the opening of the new conversation along with the files that still matter.

SignalMoveWhy
Answers slowing; “organising its thoughts” appearsSummarise and restartThe window is filling; every turn now carries the whole history
An early instruction is being ignoredRestate it, or restart with it up frontIt may now sit inside a summary rather than verbatim
Usage limit approaching mid-threadStart a new conversationLong conversations consume more; this is the help centre’s own advice
The same documents every weekPersist in a ProjectRetrieval loads the relevant parts instead of the whole pile
“Claude keeps forgetting I work in procurement”Memory, not a longer promptStable facts about you are exactly what memory stores
A single enormous first messageSplit the task, or use a ProjectEdge cases like this can still hit the limit

Memory is not the context window

Memory is a separate mechanism, and the exam likes the confusion. Claude generates and stores memory as a set of individual topics as you chat, rather than summarising conversations after they end, and it holds things like your role, your professional context, your ongoing work and your communication preferences. It is on by default for Free, Pro and Max plans; Team and Enterprise plans have it off by default and an owner can enable it. Each project has its own separate memory space and dedicated project summary, which keeps contexts from bleeding into each other.

You can see and manage it. Memory entries appear under Topics in the memory settings, where they can be viewed and edited, and you can simply tell Claude in chat to remember something. Pausing keeps what exists but stops new entries being created; resetting permanently deletes everything and cannot be undone. By default sensitive categories — health, politics, religion and similar — are not remembered unless you opt in, and some things are never saved at all, including government identification numbers, financial account numbers and criminal history. Incognito chats sit outside all of it: they are not saved to history or memory and are not referenced later.

Chat search is the third mechanism and is different again: on paid plans Claude can search and reference your previous conversations, with chats outside projects searched separately from the chats inside a given project, and it can be switched off in the memory settings. Memory tells Claude who you are; chat search finds what you said; the context window is what is in front of it right now.

Four places context can live

In front of the model now

  • Context window — this conversation, its attachments, the Project instructions
  • Fixed size; fills up; partly reserved for the reply
  • Cleared completely when you start a new chat

Stored for later

  • Memory — topics about you, per project, viewable and editable
  • Project knowledge — your documents, retrieved when relevant
  • Chat search — your past conversations, on paid plans
Ask which of these a scenario is really about. Almost every wrong answer in this objective is the right advice about the wrong one of these four.

A closing caution. None of these mechanisms makes an answer true, and a summary is itself an output that can lose or distort something. Check a handoff note before you build a second day of work on it, the same way you would check any other output — which is the discipline of Domain 2.

Traps the wrong answers are built from

Tempting but wrongDo this instead
Restarting a long conversation with nothing carried forwardAsk for a handoff note of decisions, constraints and open questions, check it, then restart.
Treating memory as a place to store documentsMemory holds topics about you; documents belong in a Project knowledge base.
Repeating a standing instruction in every messagePut it in the Project instructions so every conversation inherits it.
Assuming an instruction from sixty messages ago is still verbatimRestate constraints at the point where they matter in a long thread.
Pushing on in a thread that is slowing and eating the allowanceSummarise and start a new conversation — the documented advice for a long chat near the limit.

You should now be able to

  • Explain what occupies the context window and why part of it is reserved.
  • Recognise the signs that a conversation is approaching its limit and what the app does about it.
  • Choose between restarting, summarising and persisting for a given situation.
  • Write a handoff note that carries decisions, constraints, current state and open questions.
  • Distinguish the context window from memory, project knowledge and chat search.
  • Describe how memory is scoped, controlled and limited, including per-project memory and incognito chats.

Practice questions

Original questions written for this lesson, in the exam’s style. Answer first, then open the reasoning — every option is explained, including why the wrong ones are tempting.

  1. Question 1

    A consultant has worked for two days in a single conversation on a client strategy. Responses have slowed, she has seen Claude “organising its thoughts”, and a formatting rule she set on the first morning is no longer being followed.

    What is the best course of action?

    1. ARepeat the formatting rule and continue in the same conversation indefinitely.
    2. BStart a brand new conversation and re-describe the work from memory.
    3. CRequest a handoff note of decisions, constraints and open questions, then restart with it.
    4. DSwitch to a model with a larger context window and carry on.
    Show answer and reasoning
    1. AIncorrect. It fixes today’s symptom but leaves an increasingly expensive thread that will drift again.
    2. BIncorrect. Restarting is right; doing it without a handoff note discards two days of decisions.
    3. CCorrect. The window is filling and history is being summarised; a checked handoff note carries forward what matters into a fresh, cheaper thread.
    4. DIncorrect. It postpones the problem without addressing the drift, and the thread keeps costing more each turn.

  2. Question 2

    A team lead on a Team plan says Claude never remembers that he works in regulatory affairs, even though a colleague on a Pro plan says Claude remembers her role automatically.

    What is the most likely explanation?

    1. AMemory is off by default on Team plans and an owner needs to enable it.
    2. BMemory only records facts stated more than three times.
    3. CHis conversations are too short for memory to form.
    4. DMemory works only inside projects, never in ordinary chats.
    Show answer and reasoning
    1. ACorrect. Memory is documented as on by default for Free, Pro and Max, and off by default for Team and Enterprise until enabled.
    2. BIncorrect. No such threshold is documented; memory builds topics as you chat rather than by repetition count.
    3. CIncorrect. Length is not the stated condition; plan-level availability is.
    4. DIncorrect. Projects have their own memory space, but memory is not restricted to projects.

  3. Question 3

    An analyst uploads a 300-page regulatory filing as the very first message of a new conversation and asks eight detailed questions about it. The conversation immediately struggles.

    Which TWO responses are most appropriate? (Select 2.)

    1. APut the filing into a Project knowledge base so relevant sections are retrieved rather than all loaded.
    2. BSplit the work into focused conversations, each covering a defined part of the filing.
    3. CRely on automatic context management to handle it, since summarising is automatic.
    4. DAsk all eight questions in the same message to save context.
    5. ETurn memory on so the filing is remembered for later.
    6. FSwitch to the fastest model, which handles large documents more efficiently.
    Show answer and reasoning
    1. ACorrect. Projects use retrieval so the model works with larger volumes efficiently, and capacity expands as the limit approaches on paid plans.
    2. BCorrect. A very large first message is a documented edge case that can still hit the limit; smaller scoped turns avoid it.
    3. CIncorrect. Automatic summarisation helps as a conversation grows, but a single oversized first message is exactly the edge case it may not cover.
    4. DIncorrect. The questions are a tiny fraction of the load; the filing is the problem, not the questions.
    5. EIncorrect. Memory stores topics about the user, not uploaded documents.
    6. FIncorrect. The fastest tier documents the smallest context window, which makes this worse rather than better.

  4. Question 4

    A manager wants Claude to stop referencing an old project she no longer works on, which keeps appearing in its answers as context about her role.

    What should she do?

    1. AStart using incognito chats for all her work from now on.
    2. BReset memory entirely, since individual entries cannot be changed.
    3. CAsk Claude in every conversation to ignore the old project.
    4. DOpen the memory settings, find the entry under Topics, and edit or remove it.
    Show answer and reasoning
    1. AIncorrect. Incognito prevents new saving but does nothing about the entry that already exists.
    2. BIncorrect. Reset is permanent and irreversible, and entries can in fact be viewed and edited individually.
    3. CIncorrect. A per-conversation instruction is a workaround that has to be repeated forever.
    4. DCorrect. Memory entries are viewable and editable under Topics, which is the targeted fix and leaves everything else intact.

Practise every question in Domain 3 →

Sources

Drafted with AI assistance and checked against the sources above; expert review is in progress. Spotted an error? Tell us and it gets fixed, dated and listed on how this is written.