Task statement 3.4 · Product and model selection

3.4 · Context Limits and Memory

In Wrenfield Housing, the course example, the communications lead drafts the annual tenant report (1.2) in one chat over three weeks. By week three, Claude's replies ignore the style decisions agreed in week one, and each reply is slow, because the chat now carries three weeks of drafts.

Why it's tested

The tempting fix for a drifted chat is to stay in it and repeat the rule. Each reminder only adds one more message to a conversation that is already too long, so the next reply drifts again. The opposite habit fails too, because a blank chat with the brief retyped from memory loses the decisions and corrected figures that nobody wrote down.

KnowledgeWhat fills a chat, and what happens as it grows

Context is what Claude can see when it answers: your prompt, the conversation so far and any attached files. The context window is the limit on how much of that Claude can work with at once, measured in tokens, the small chunks of text Claude counts. Everything in a chat fills it: messages and replies, attached files, a Project's instructions and knowledge, and the results of tools such as web search and connectors. On paid plans even the smallest window holds 200,000 tokens, about 500 pages of text (checked 2026-10-15), so a long chat rarely hits a hard wall. Its quality slips first: Anthropic's documentation says accuracy and recall degrade as the token count grows, and calls this context rot. A style rule from the fifth message ends up as a few lines among hundreds of pages, so replies drift from it. This rules out any answer that treats a bigger window or a firmer reminder as the cure for a drifted chat.

As a single chat grows, five things happen in this order:

  1. Everything stays: Claude works from the whole conversation so far each time it answers.
  2. Material adds up: files, pasted tables and tool results sit alongside the messages.
  3. Early instructions get buried: a first-day decision becomes a small part of a large context.
  4. Claude summarises near the limit: with code execution, a feature that lets Claude run code, turned on, Claude summarises earlier messages and may say it is organising its thoughts.
  5. Usage climbs: longer chats use more of your usage limit, and at a length limit the Help Center's answer is a new chat.

Interactive · one report, three weeks, three moves

Four stages of the annual report work, from top to bottom. On the left is what each chat holds; on the right is how Claude's replies treat the three style decisions. In week 1, chat A holds the brief and the decisions, and replies follow them. In week 3, the same chat holds about 90 messages, and replies drift from the decisions. A restart opens chat B with a checked summary, and replies follow the decisions again. Persisting puts the decisions in the Tenant comms Project's instructions, so every chat there starts from them. A drifted reply is dotted and labelled drifted; a stage not reached yet is dashed and labelled not yet. WHAT THE CHAT HOLDS CLAUDE’S REPLIES · THE 3 STYLE DECISIONS Week 1 · chat A brief, sources, last year’s report + 3 style decisions agreed GIVES FOLLOWED “tenants”, short sentences, every figure dated Week 3 · still chat A about 90 messages: 8 sections, drafts, revisions, tables NOT YET GIVES DRIFTED “customers”, 30-word sentences → SLOW; DECISIONS BURIED NOT YET Restart · new chat B checked summary: 3 decisions, 71% satisfaction, open items NOT YET GIVES FOLLOWED AGAIN same 3 decisions, far less for Claude to weigh NOT YET Persist · Tenant comms Project 3 decisions written into the Project’s instructions NOT YET GIVES STARTS FROM THEM every chat in the Project, next year’s report included NOT YET A long chat buries early decisions; a checked summary or the Project carries them forward.

Step through three weeks of the annual report and see where the style decisions survive.

Week 1: chat A holds the brief, the sources and three agreed style decisions, so Claude’s replies follow all three.

Week 3, same chat: about 90 messages now surround the decisions. Replies say “customers”, run to 30-word sentences and come slowly.

Restart: the lead checks a summary of the decisions, figures and open items, and opens chat B with it. Replies follow the decisions again.

Persist: the three decisions go into the Tenant comms Project’s instructions, so every chat there, including next year’s report, starts from them.

Dotted and labelled DRIFTED: replies ignore the decisions. Dashed and labelled NOT YET: a stage still to come.

KnowledgeRestart, summarize or persist

Knowing when to restart, summarize or persist comes down to what must cross into the next chat. To restart is to begin again in a new chat from a fresh prompt. As 1.3 showed, a new chat carries only what you write into it. To summarize is to ask Claude for the decisions, facts and open items in the long chat, check them against your sources, and open the new chat with them. To persist is to put what must survive every chat somewhere lasting. A Project's instructions apply to every chat in it, and on a Team plan the Project can be shared with colleagues. Chats in a Project do not share their context, though, unless it is added to the Project's knowledge:

Restart

The chat has drifted, or the task has changed, and the brief fits in a fresh prompt. Nothing else from the old chat is needed.

Carries forwardOnly the new prompt.

Summarize

The same piece of work continues, and the long chat holds decisions and corrected facts it still needs.

Carries forwardA summary you have checked, pasted into the new chat.

Persist

A rule or fact must hold in every future chat, or for colleagues too.

Carries forwardProject instructions or knowledge, or Claude's memory for your own preferences.

Claude's memory feature is the other lasting place. Memory saves topics about your work and preferences from your chats, on its own or when you ask, and each Project keeps its own memory. On Team and Enterprise plans it is off for each member until the owner makes it available and the member turns it on. Settings > Memory lists every topic it holds, to edit or delete, and lets you pause or reset it. The "+" menu turns it off for one new chat. Memory belongs to one member's account, and even owners cannot read it, so it suits your own working preferences, not a team's rules. This rules out any answer that relies on one person's memory to hold a rule every colleague must follow.

SkillWorked example: the report chat that forgot its style

In week one the lead and Claude agreed three style decisions: say "tenants", never "customers"; keep sentences to 20 words or fewer; and give every figure its period and source, such as "71% (2025–26 survey)". By week three, with about 90 messages in the chat, a reminder did not hold. The fix was a checked summary in a new chat:

# week 3, same chat
Remember: say "tenants", never "customers".
→ next section: "customers" twice, 30-word sentences
# summarize, in the old chat
List the style decisions we agreed, the checked figures, and the open items.
→ satisfaction with repairs: 81%   (the survey says 71%)
# restart: new chat, summary corrected to 71%, then the next task
→ "tenants" throughout, short sentences, every figure dated

The cause was the length of the chat, not a weak model or a vague rule. The decisions were still in the chat, but they sat among three weeks of drafts, so a reminder became one more message in the crowd. The summary carried the decisions forward, and checking it mattered. It repeated the 81% from the first one-message draft, corrected in week one, because a long chat still holds its old mistakes. Last, the lead persisted the three decisions in the instructions of the team's Tenant comms Project (setting one up is 5.1), so next year's report starts from them. The lead left them out of personal memory, because colleagues draft in that Project too.

Automatic summarising is not a fresh start. It needs code execution turned on, and long chats that trigger it use more of your usage limit. Your full history is kept, so the old drafts stay part of the chat.

Ask one question: what from this chat must the next one have?
Nothing beyond a brief you can write freshRestart: a new chat with a clean prompt.
Decisions and facts this piece of work still needsSummarize: ask for a summary, check it, and open the new chat with it.
Rules every future chat, or every colleague, must followPersist: Project instructions or knowledge; memory only for your own preferences.
Takeaway
  • Messages, files, Project knowledge and tool results all fill the context window, and an early instruction gets buried as a chat grows.
  • A reminder in a drifted chat adds to the problem; a new chat with only what it needs fixes it.
  • Check a summary before you carry it over, because it can bring back a mistake the long chat once held.

check yourself

0 / 3 answered

A sales coordinator has planned a 250-guest conference in one chat for two weeks, and the chat now holds over 150 messages. Recent replies suggest venues that are too small and forget the agreed catering budget. The coordinator must carry on tomorrow, starting from every decision already agreed.What is the most effective next step?

Select one.

A finance team lead on a Team plan wants every chat their six colleagues have about the monthly cost report to follow two rules: amounts shown in thousands, and only the cost-centre names from an approved 40-row list. The rules must apply without anyone pasting them in.Which TWO steps best achieve this?

Select two.

A consultant uses Claude on their firm's Team plan. In a chat outside any Project, they wrote: "Remember that my client reports use British spelling." The next day, a new chat used American spelling. Under Settings > Memory, no topics are listed at all, although the consultant has used Claude daily for a month.What is the most likely root cause?

Select one.

sources

  1. Claude Help Center, "How do usage and length limits work?": the context window as Claude's working memory for a chat; tools and connectors are token-intensive; automatic context management summarises earlier messages near the limit and needs code execution; Claude may show it is organising its thoughts; longer conversations that trigger it use more of your usage limit; at a length limit, start a new conversation. Checked 2026-10-15.
  2. Claude Help Center, "How large is the context window on paid Claude plans?": 200K tokens is about 500 pages of text; newer models have larger windows; automatic context management on paid plans with code execution enabled. Checked 2026-10-15.
  3. Claude Platform Docs, "Context windows": the context window is all the text Claude can reference when it responds; previous turns accumulate in it; as token count grows, accuracy and recall degrade (context rot). Checked 2026-10-15.
  4. Claude Help Center, "Use Claude's chat search and memory to build on previous context": what memory saves; off for each Team and Enterprise member until turned on; owners enable it and cannot view members' memories; each Project has its own memory; Settings > Memory to view, edit or delete topics; pause, reset, and memory off for one chat. Checked 2026-10-15.
  5. Claude Help Center, "How can I create and manage projects?": project instructions apply to all chats in the Project; context is not shared across chats unless added to the project knowledge. Checked 2026-10-15.
  6. Claude Help Center, "What are projects?": on Team and Enterprise plans, Projects can be shared with other members of the organisation. Checked 2026-10-15.
  7. Wrenfield's three-week report chat, its three style decisions and the Tenant comms Project's instructions are this course's example.