When to restart, summarize or persist

A compliance coordinator has been rewriting the travel expenses policy with Claude in the same chat for three weeks. At the start they agreed on rules: numbered clauses, a definitions section, nothing about the old reimbursement rates. The chat now holds dozens of drafts, pasted documents and side discussions. In the last few replies, Claude stopped numbering and once quoted an old rate. Reminding it of the rules works for one or two messages and then the drift returns.

It isn’t that Claude “got tired.” It’s a context problem.

What the context window is

The context window is how much text Claude can take into account at once: your messages, its replies, the files. It’s large, but not infinite. In a very long conversation full of drafts, the instructions from the beginning end up buried under thousands of later lines and lose weight.

On paid plans, when a conversation nears the limit, claude.ai can automatically summarize earlier messages to keep going (code execution has to be enabled). That avoids interruptions, but it doesn’t tidy a chat that mixes three weeks of versions and topics: what was already confusing stays that way.

The three moves

Move	When	How
Summarize and restart	The chat is very long and Claude starts forgetting what was agreed	Ask for a summary of the rules and the state of the work; open a new conversation that starts with that summary
Plain restart	You’re changing topic or what came before isn’t useful	New chat, carrying nothing over
Persist context	Something applies across many conversations	Move it into a Project or memory instead of repeating it

For the coordinator, the move is the first one: “Summarize the drafting rules we agreed on, the finished sections and what’s left.” She checks the summary (that the old rate didn’t sneak in), opens a new chat with it and carries on with her four remaining sections. If the rules will be useful for other policies, they belong in a Project’s instructions.

Memory and Projects

claude.ai’s memory stores topics from your conversations so you don’t have to repeat preferences; you can view, edit or pause it in settings.

The key point for the exam is project memory: each Project has its own memory, separate from other Projects. Inside a Project, Claude uses that Project’s memory and the memory from your chats outside Projects, but not memory from other Projects. And if you want to isolate it completely, you can turn off the use of account memory in that Project.

An HR consultant serves a law firm (formal tone) and a gaming company (its own jargon). She works in loose chats with memory on, and one client’s jargon started showing up in letters to the other. The solution: one Project per client, with continuity for each and no mixing.

An incognito chat isn’t saved to history or memory. It’s for a query you don’t want kept, not for separating clients: it loses exactly the continuity she wants.

Common traps

• Repeating all the rules in every message of an endless chat. A patch that doesn’t last.
• Switching to the most capable model in the same chat so it “remembers more.” It doesn’t tidy the conversation.
• Staying in the same conversation forever and fixing the drift by hand at the end.
• Putting two clients in one long chat “so it can see which client each request is for.”
• Telling it to “forget the other client” at the start of each chat, instead of separating the spaces.
• Using incognito for everything, losing useful memory.

On the exam

“Weeks in the same chat,” “dozens of drafts,” “reminding it works for a while” point to summarizing and continuing in a new conversation. “The clients get mixed up” or “context from one job showing up in another” point to separating into Projects, because of their own memory. “Repeats the same thing in every chat” points to persisting in a Project. Rule out options that add repetition, switch models or rely on asking Claude to forget.