What you will be able to do
- Explain why a long chat starts ignoring decisions made early in it
- Restart drifting work in a new chat with a decisions list, and say why the alternatives fail
- Choose between in-chat context, Project knowledge and memory for work that spans many sessions
- Recognize the signs that a long chat needs a restart rather than a better prompt
- Review memory when long work restarts, and pick the edit or deletion that removes a stale entry without losing the rest
Key concept
Context window — The fixed amount of text Claude can pay attention to in one conversation. Anything outside it has no effect on the answer, and when a chat outgrows it, quality drops off suddenly instead of gradually.
1.Why long chats lose early decisions
Everything Claude can use in a chat has to fit inside its context window, a fixed space it can attend to all at once. Claude's help center calls the window Claude's working memory: it limits how much content Claude can process and keep in view at one time. A rule agreed in the first hour of a three-hour drafting session is still in the transcript. Being in the transcript does not mean Claude is still attending to it.
The limit acts like a cliff, not a slope. Work goes well while the material fits, and then it stops working. The usual failure is silent truncation, and you won't always get a warning. You can predict where trouble starts: very long conversations, very long documents, critical details buried in the middle of long input, and expecting one session to carry over into the next. Window size depends on the model. On paid plans the newest models go up to 1M tokens and others to 500K or 200K, and part of the window is always reserved for Claude's reply.
Corrections are not learning. A fix you typed an hour ago is one more message competing for space with everything after it, so in an overloaded chat the same drift keeps coming back. Knowing how the window works tells you how to arrange context, what to put first, and when to start a fresh chat.
Restart when the symptoms point at the window rather than at your prompt. Three signs mean a fresh chat will do more than another rewrite:
- A correction holds for a turn or two, then the same drift returns.
- Claude ignores a decision settled early in the session, with no warning that anything was dropped.
- The work is moving into a new phase, say from outlining to drafting, which is a natural point to split long work into chunks.
Rewording fixes an unclear request. It cannot make room in a full window, so once these signs appear, a restart beats a fourth attempt at the same instruction.
Checkpoint 1 of 5· Check yourself
You are three hours into a drafting session. Claude's answers were steady. Then, within a few exchanges, it starts ignoring a formatting rule you agreed on in the first hour. No warning appeared. What best explains this?
The context window is a fixed size, so quality holds while the material fits and then drops suddenly. The usual failure is silent truncation: the rule is still in the transcript, but Claude is no longer attending to it.
“This property has a cliff rather than a gradient. Silent truncation is the failure mode, and you won't always be warned.”Source: academy.claude.com
2.Restarting with a decisions list
When a long chat starts ignoring settled decisions and you still need to keep working, the strongest move is a handoff. Ask Claude to list every decision, constraint and open item agreed so far, and check the list yourself. Then open a new chat and paste the list as your first message. The new chat starts with the critical material at the front of an almost empty window. That applies the strategies the limit calls for: put important material first, split long work into chunks, and supply critical context again.
Three alternatives are tempting, and each one aims at the wrong problem:
- Retyping the rules at the end of every prompt in the same chat, even in capitals. The window is still overfull, and emphasis adds no capacity. You are resupplying context into the same place where it is getting lost.
- Deleting early messages through memory settings. Memory settings control what Claude carries between chats. They do not trim the messages inside a conversation.
- Switching the chat to incognito. You choose incognito when you start a chat, and it controls whether the chat is saved. An existing chat can't be converted, and incognito does nothing about context size.
Claude can also manage length on its own, with limits. With code execution enabled, Claude summarizes earlier messages when a conversation approaches the context window limit, and you may see it "organizing its thoughts". That keeps the chat going, but you don't control what the summary keeps. Conversations that trigger this summarizing also use more of your usage limit, so if you are getting close to that limit in a long chat, starting a new conversation is the better choice. A decisions list you have checked yourself is a summary you can trust.
Checkpoint 2 of 5· Check yourself
A colleague has to finish a report today. Their chat is now very long, and Claude keeps ignoring constraints they agreed on early in it. What should they do next?
A checked decisions list pasted into a new chat puts the critical material first in an almost empty window. Retyping rules in the overloaded chat adds no capacity. Memory settings don't trim messages inside a chat, and an existing chat can't be switched to incognito, which controls saving, not context size.
“Apply context-as-leverage strategies: front-loading important material, chunking long work, re-supplying critical context”Source: academy.claude.com
Checkpoint 3 of 5· Exam question
Priya, a marketing manager, has spent three hours in a single Claude chat drafting a campaign. Early in the chat the team agreed on a playful tone and a 120-word limit, but Claude's latest drafts now ignore both. She needs to keep going today. What is the best next step?
Correct answer: C — Ask Claude to list the agreed decisions, then open a new chat and paste that list as the first message.
- A. Incorrect: memory settings store distilled facts about you, not the messages of a live chat, so nothing there can be trimmed to free up room.
- B. Incorrect: repeating the rules in capitals leaves the bloated history in place, and the drift that caused the problem continues.
- C. Correct: a long chat can drift from early instructions, so summarizing decisions and restarting with that summary gives Claude a short, clean context that contains the rules.
- D. Incorrect: incognito only controls whether a chat is saved to memory and history; it does not remove earlier messages from the current conversation.
3.Choosing between the chat, a Project and memory
A handoff note rescues one session. Work that runs across many sessions needs a place every new chat can start from. Claude has several such places, and each one reaches a different set of chats.
| Where context lives | Reach | Best for |
|---|---|---|
| In-chat context | This conversation only, up to the context window limit | Task-specific instructions and the draft you are working on |
| Project knowledge and instructions | Every chat inside that project; project instructions apply only to chats in it | Reference files, rules and briefs that a body of work depends on |
| Memory | Saved as topics while you chat; a project uses its own memory plus your account memory, never another project's | Your role, preferences and ongoing work that should follow you |
| Search past chats (paid plans) | Finds details in earlier chats when you ask; inside a project, only that project's chats | Pulling up something specific you discussed before |
Project knowledge is the deliberate choice for material that work depends on. Files uploaded to a project's knowledge base give every chat in that project the same background. On paid plans, projects use retrieval, loading only the relevant content into the context window, so they can hold far more than a single chat. Keep project instructions short and general, and save task-specific direction for the chat itself.
Memory takes the least effort, and its scope is what catches people out. Say a pricing rule was settled in a Project called Pricing Review and is now needed in a Product Launch Project. Claude inside Product Launch sees that project's memory and your account memory, but never Pricing Review's memory. Searching past chats from inside a project also stays within that project. The reliable fix is to put the rule where the work happens: add it to Product Launch's knowledge or instructions.
Everything in project knowledge can influence every answer. If an outdated 2025 price sheet sits next to its 2026 replacement, both are available as background, and quotes can mix figures from the two. The help center's advice is to clear out files you no longer use. Keeping project configuration current over time is covered in configuration lifecycle.
Checkpoint 4 of 5· Match them up
Match each place where context can live to how far it reaches and what it is best used for.
Tap a term, then the definition that fits it.
Each place reaches a different set of chats. In-chat context ends with the conversation. Project knowledge covers every chat in that project. Memory follows you, but inside a project it never draws on another project's memory. Search past chats run from a project stays inside that project.
“While working in a project, Claude uses the project's memory alongside your memory from chats outside of projects, but not memory from other projects.”Source: support.claude.com
4.Reviewing memory when long work restarts
A new chat clears the context window but not memory. Whatever memory picked up during a long stretch of work comes along into the restart, including details that were true on Monday and wrong by Friday, such as a launch date that has since moved. A stale message drops out of view when you start fresh; a stale memory entry keeps shaping every chat after it. Memory quality therefore needs its own review, separate from the handoff.
The restart is the best moment for that review, because you have just checked a decisions list and know exactly what changed. Settings > Your Claude, then Manage next to Your memory, shows memory as Claude sees it; read it against the list. An entry that is out of date calls for an edit through the usual correction routes. An entry that should not be there at all calls for a deletion, and the deletion controls differ sharply in how much they remove:
- Deleting an individual memory removes that one entry and leaves the rest in place.
- Pause memory deletes nothing. Claude keeps its existing memory but stops using it and stops making new memories, so the wrong entry is still there when you unpause.
- Reset memory permanently deletes all memories, project memories included, and can't be undone.
- Deleting a project deletes that project's memory along with it.
Deleting the chat is not the same as deleting the memory. On Team and Enterprise plans, memory entries generated from a deleted conversation aren't removed, so clearing out the drifting chat leaves its mistakes in memory. Delete the entry itself, then glance at the panel again before the next long session begins.
Checkpoint 5 of 5· Check yourself
Your team plan's memory still holds a launch date that moved two weeks ago. You are restarting the work in a new chat and want that one entry gone while everything else Claude remembers stays. Which action does that?
Deleting the individual memory removes only that entry. Deleting the chat leaves entries generated from it in place, pausing keeps existing memory while stopping all of it from being used, and reset permanently deletes everything.
“memory entries generated from it aren’t removed, but members can delete individual memories while memory is on.”Source: support.claude.com
Sources
(7)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Think of the context window as Claude's working memory that determines how much content it can process and remember at once.”
↩︎ Why long chats lose early decisions“On paid plans, the newest models support up to a 1M token context window, while others support 500K or 200K tokens.”
↩︎ Why long chats lose early decisions“When your conversation approaches the context window limit, Claude summarizes earlier messages to continue the conversation seamlessly.”
↩︎ Restarting with a decisions list“Code execution must be enabled for automatic context management.”
↩︎ Restarting with a decisions list“Longer conversations that trigger automatic context management consume more of your usage limit.”
↩︎ Restarting with a decisions list“Try starting a new conversation if you're approaching your usage limit in a longer chat.”
↩︎ Restarting with a decisions list“by only loading relevant content into the context window.”
↩︎ Choosing between the chat, a Project and memory“Reserve task-specific instructions for the chat itself.”
↩︎ Choosing between the chat, a Project and memory“Remove unused project files: Regularly clean up files you're no longer actively using in your projects.”
↩︎ Choosing between the chat, a Project and memory - 2.
“This property has a cliff rather than a gradient.”
↩︎ Why long chats lose early decisions“Silent truncation is the failure mode, and you won't always be warned.”
↩︎ Why long chats lose early decisions“Limitation zone: very long documents or conversations, expecting continuity across sessions, burying critical info in the middle of long input.”
↩︎ Why long chats lose early decisions“Knowing how the window works tells you how to structure context, when to front-load, and when to start fresh.”
↩︎ Why long chats lose early decisions“The model doesn't learn from your corrections. It only responds to what's currently in context.”
↩︎ Why long chats lose early decisions“Apply context-as-leverage strategies: front-loading important material, chunking long work, re-supplying critical context”
↩︎ Why long chats lose early decisions“Apply context-as-leverage strategies: front-loading important material, chunking long work, re-supplying critical context”
↩︎ Restarting with a decisions list“Everything the AI is paying attention to lives inside a fixed-size workspace called the context window.”
↩︎ Key concept“The model doesn't learn from your corrections. It only responds to what's currently in context.”
↩︎ Exam trap 1“This property has a cliff rather than a gradient. Silent truncation is the failure mode, and you won't always be warned.”
↩︎ Checkpoint - 3.
“once you start an incognito chat, it cannot be converted to a regular chat or saved to your history.”
↩︎ Restarting with a decisions list“once you start an incognito chat, it cannot be converted to a regular chat or saved to your history.”
↩︎ Exam trap 2 - 4.
“You can upload relevant documents, text, code, or other files to a project's knowledge base”
↩︎ Choosing between the chat, a Project and memory - 5.https://support.claude.com/en/articles/10185728-understanding-claude-s-personalization-featuresOfficial docs
“These instructions only apply to chats within that project.”
↩︎ Choosing between the chat, a Project and memory - 6.https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-contextOfficial docs
“Claude uses the project's memory alongside your memory from chats outside of projects, but not memory from other projects.”
↩︎ Choosing between the chat, a Project and memory“Individual project conversations (searches are limited to within each specific project).”
↩︎ Choosing between the chat, a Project and memory“Claude keeps its existing memory, including sensitive topics if you have it turned on, but won't use memory or make new memories.”
↩︎ Reviewing memory when long work restarts“Reset memory: Permanently deletes all memories including project memories.”
↩︎ Reviewing memory when long work restarts“Project memory is deleted when the project is deleted”
↩︎ Reviewing memory when long work restarts“memory entries generated from it aren’t removed, but members can delete individual memories while memory is on.”
↩︎ Reviewing memory when long work restarts“Claude uses the project's memory alongside your memory from chats outside of projects, but not memory from other projects.”
↩︎ Exam trap 3“Reset memory: Permanently deletes all memories including project memories.”
↩︎ Exam trap 4“While working in a project, Claude uses the project's memory alongside your memory from chats outside of projects, but not memory from other projects.”
↩︎ Checkpoint - 7.https://support.claude.com/en/articles/12123587-import-and-export-your-memory-from-claudeOfficial docs
“To view your memory exactly how Claude sees it, go to Settings > Your Claude”
↩︎ Reviewing memory when long work restarts