Home › Study guides › CCAO-F › Domain 3 › Lesson 3.4
CCAO-F · Domain 3 · 12% of the exam · Lesson 3.4 · 18 min read
Context and memory: when to restart, summarise or persist
Why long Claude chats drift, the signs it is time to act, and when to restart, hand off a reviewed summary, or persist context in a Project or memory.
Written against objective 3.4 of the official CCAO-F exam guide (Version 1.0, effective July 2026). An independent resource, not affiliated with Anthropic; the practice questions are written from scratch.
3.4.1 Why a six-week chat starts getting the facts wrong
Declan is a management consultant, six weeks into a pricing review for Maplewick Home, a homeware retailer with 60 stores. He has run the whole engagement in one Claude chat. It now holds notes from twenty interviews, three versions of his pricing recommendation and two sets of margin figures, the second issued after finance corrected its supplier costs. At the kickoff, the client's chief executive set one firm rule: the recommendations must not involve closing stores.
On Friday Declan asks the chat for a one-page board summary of version 3. It reads well, but he checks the numbers anyway. The margin on the premium range says 41%, the version-2 figure; finance's corrected figure is 38%. And the closing paragraph suggests "consolidating two underperforming stores", the one move the client ruled out.
Claude hasn't broken down. It is answering from the conversation, which holds all three versions side by side with nothing to say which one is current. The chief executive's rule is in there too, under five weeks of later material. Claude only knows what is in front of it, and a long chat puts a great deal in front of it, much of it out of date.
The mechanism behind this is the context window: everything Claude can consider at once in one conversation, and much of what goes into it is up to you. Managing it comes down to three moves: RESTART the conversation, SUMMARISE what is worth keeping into a fresh one, and PERSIST what every future chat needs in a place built for it.
Three moves for managing context
RESTART
SUMMARISE
PERSIST
3.4.2 What Claude works from: the context window
Why would Claude quote a figure Declan replaced a fortnight ago? Start with what the context window holds: your messages, Claude's replies and any files you have attached. The Help Center describes it as Claude's working memory. Don't confuse it with the memory feature later in this lesson, which carries things between chats.
Each time you send a message, Claude processes it together with the conversation so far, not on its own. That is what lets you say "shorten the second paragraph" without pasting the paragraph again. It also means nothing drops out just because you replaced it: the version-2 margins stay in Declan's chat after version 3 arrives.
What Claude reads when Declan asks for the board summary
The window's size depends on the model and plan. On paid plans, the Help Center gives 200,000, 500,000 or 1 million tokens depending on the model, a token being a small chunk of text. Size is rarely the first problem, but a long chat costs you twice. It draws more on your usage limits, which cap how much you can use Claude over a period, and it moves toward the length limit, which caps how long one conversation can grow.
When code execution is switched on (the feature that lets Claude run code and create files), Claude can summarise earlier messages as a chat nears the length limit, keeping the full history for reference. This uses more of your usage. It keeps the chat going, but it can't know which version you approved, so sorting current facts from superseded ones is still your job.
3.4.3 When a long chat works against you
If the window can hold six weeks of work, why did the summary go wrong? Because room is not the same as clarity. Anthropic's documentation on context windows puts it plainly: more context isn't automatically better, and as it grows, accuracy and recall degrade. What is in the conversation matters as much as how much fits.
Early details, such as the kickoff rule, get buried under weeks of later material, while old versions, abandoned assumptions and rejected ideas never leave. Picture a project binder in which every page is filed by date and nothing is ever stamped "superseded". Ask a capable colleague for a board summary from it, and they may quote the version-2 page, because it looks just as official. Superseded content stays in the conversation in the same way, and Claude can draw on it as readily as on the current facts.
Declan's first instinct is to type "No, use 38%" and carry on. It works for a reply or two, then 41% is back. A correction can only add to the conversation; it can't take the old figure out. Anthropic's Claude 101 course makes the practical point: when a conversation has gone off track, it is sometimes faster to open a new chat with a clearer prompt than to keep redirecting.
| Symptom | In Declan's chat | What it tells you |
|---|---|---|
| Mixes old and new versions | The version-3 summary quotes version-2 margins | Superseded material is competing with the current facts |
| Forgets an early constraint | It proposes consolidating two stores | The rule is buried under later material |
| Repeats a figure you corrected | 41% returns two replies after "use 38%" | Corrections add to the pile; they don't replace anything |
| The thread has changed topic | He has started next quarter's proposal in the same chat | Two jobs now share one context |
Memorise the four symptoms: each is a signal to act on the context itself, not to keep correcting inside the thread.
3.4.4 Restart, or summarise and then restart
Once you have spotted a symptom, the question is how much of the conversation to carry forward.
RESTART means opening a new chat and briefing it from scratch. It fits when the topic has changed, like Declan's proposal for next quarter, or when a muddled thread holds nothing worth keeping. SUMMARISE fits when the thread holds work you need: decisions, rules, approved figures. You ask Claude for a handoff summary, check it, and start a new chat with the corrected version. It works like a shift handover on a hospital ward: the outgoing nurse passes on current status, recent changes and what to watch, not the whole twelve hours.
Here is the prompt Declan uses. The second and third lines ask for current facts only and say what to do when Claude isn't sure which version holds.
We are going to continue this work in a new chat. Write a handoff summary I can paste in to start it.
Include only what is current: the client's goal, every rule the client has set, the decisions we have made, and the latest figures with the version they come from.
Leave out superseded versions, rejected options and replaced figures. Where the chat holds two versions of a figure, give only the latest. If you are not sure whether something is still current, write [CHECK] instead of guessing.
List open questions separately at the end.
Keep it under 400 words, in short labelled sections.
Then comes the step people skip: the review. A summary of a muddled conversation can carry the muddle forward, and it reads just as confidently either way; confidence is not evidence. Declan checks it against the version-3 spreadsheet and his kickoff notes. The premium margin is right, but the entry-range margin is still the version-2 figure, and "store consolidation" appears as an open question. He corrects both, gives the no-closures rule its own line, and starts a new chat with the reviewed summary and the version-3 figures attached. The board summary comes back clean.
3.4.5 Persist what every future chat needs
A reviewed summary fixes this week, but the no-closures rule has to be in every chat Declan opens from now on. Pasting it in each time works until the week he forgets. Anything every future chat needs belongs somewhere that brings it in without him.
For the engagement itself, that place is a Project: a workspace with its own chats, its own instructions that shape Claude's responses, and its own knowledge. Knowledge means the files Claude draws on as background in the chats inside that Project. Projects are available on every plan, including Free. The Help Center advises keeping Project instructions to general context, key guidelines and Claude's role, and leaving task-specific instructions for the chat itself.
That advice points to the real skill: choosing what to persist. Whatever you put in a Project comes back in every chat, so a mistake there repeats week after week instead of fading with one thread. Here is how Declan sorts what six weeks have produced.
| What Declan has | Where it belongs | Why |
|---|---|---|
| The review's goal and the no-closures rule | Project instructions | Every chat on the engagement must follow them |
| The client brief and the signed-off version-3 figures | Project knowledge | Approved reference material that every chat draws on |
| His preference for five-bullet executive summaries | Memory | His own working style, not a client rule |
| This week's what-if price scenarios | The conversation | Used once, and they will change |
| Versions 1 and 2 and the rejected options | Nowhere persistent | They would bring back the stale material he just removed |
| Interview notes with candid comments about named staff | Nowhere persistent | Sensitive, and only the agreed findings matter later |
Memorise the test behind the last column: persist only what every future chat should rely on.
3.4.6 What memory and chat search carry between chats
It is easy to assume Claude remembers either everything you have told it or nothing. Neither is right: across ordinary chats, Claude doesn't know what you said elsewhere unless a Project, memory or chat search brings it in.
Memory saves context from your chats as you go, as individual topics, and brings it into new conversations. The Help Center's examples include your role, the people and projects you work with, your preferences and working style, and ongoing work. Claude decides what to save, you can ask it to remember something, and each Project has its own separate memory. Memory is on by default on Free, Pro and Max plans; on Team and Enterprise it stays off until an owner switches it on. Treat it like a colleague's jotted reminders of how you work: useful, but not the official record.
Chat search, on paid plans, lets you ask Claude to find what you discussed before. It searches either your chats outside Projects or the chats inside one Project, and links back to the original chats. It finds what past chats say, including figures you have since replaced, such as Declan's version-2 margins, so check what it brings back.
What a new chat knows
The memory settings list everything Claude remembers as topics you can read, edit or delete; you can also ask Claude in a chat to change or forget something. Deleting a conversation does not remove memory entries made from it. So if memory holds something you would rather it didn't, delete or correct that topic directly.
An incognito chat, available on every plan outside Projects, doesn't use your existing memory and isn't saved to chat history or memory. It is still kept for a period, 30 days by default, and on Team and Enterprise plans it is included in the organisation's data exports. Incognito changes what Claude remembers, not what you may share.
3.4.7 The exam traps
Every trap here leaves stale or unwanted material where Claude will keep finding it.
- ✗ Correcting Claude reply by reply in a thread that keeps slipping. ✓ Restart from a reviewed summary. A correction adds to the conversation; it can't remove the old figure.
- ✗ Switching to a more capable model or a larger context window when a chat mixes versions. ✓ Clean the context first. A bigger window holds more of the same stale material, and no model can know which version you approved.
- ✗ Pasting Claude's handoff summary into a new chat unread. ✓ Check it against your own records. A summary of a muddled chat can carry the muddle.
- ✗ Carrying the whole old conversation forward, into a new chat or a Project. ✓ Carry forward only current decisions, rules and figures. The full transcript brings every stale version with it.
- ✗ Saving everything to a Project or memory so Claude "never forgets". ✓ Persist only approved, reusable material, and keep one-time facts, brainstorming and sensitive details out.
- ✗ Assuming that deleting a chat clears what Claude remembered from it. ✓ Review and edit memory directly in its settings.
Four tempting fixes for a stale chat, one real one
3.4.8 Put it together: reset a stale chat
You now have the whole method, and four symptoms tell you when to use it. Restart when there is nothing to keep, summarise and review when there is, and persist what every future chat needs: the work in a Project, your preferences in memory. The quickest way to trust this is to make a chat go stale on purpose.
Domains 5 and 6 build on these moves. Configuring a Project with instructions and knowledge (5.1) takes the persist move further, and keeping instructions, knowledge and memory current (5.4) stops persisted context from going stale in its turn. What data may go into a chat, a Project or memory at all is a privacy question (6.2).
Key takeaways
- ✓ The context window is everything Claude can consider in one conversation, and each new message is processed with it, so long chats draw more on your usage and approach length limits.
- ✓ More context is not automatically better: superseded versions, abandoned assumptions and early rules buried under later material all stay in play.
- ✓ Mixed versions, a forgotten constraint, a corrected figure coming back or a change of topic are signs to act on the context, not to keep correcting.
- ✓ Restart for a new topic or a thread with nothing to keep; otherwise ask for a handoff summary, review it against your records and start a new chat from it.
- ✓ Persist only approved, reusable material in a Project's instructions and knowledge, and keep one-time facts, brainstorming and sensitive details out of persistent places.
- ✓ Across chats, Claude knows only what a Project, memory or chat search brings in; review memory in its settings, and use incognito for chats that shouldn't be remembered.
Check your understanding
4 questions written for this lesson, then one from the CCAO-F question bank on the same topic. Every answer option is explained, including the ones you did not pick. Nothing is stored.
42 CCAO-F questions on Domain 3, free
Every question in the bank is tagged to a domain, so you can drill 42 questions on Product and Model Selection alone, or sit the full 60-question timed simulator.
Open the CCAO-F question bank → Back to Domain 3 →
The question bank is free. It asks for an account only because the quiz engine has to store answers to score them and show which domains are weak. The questions on this page need nothing.