As a conversation grows longer, it can feel like Claude reflects your earlier instructions or context less. This is a common pattern with language models handling long context, and a few habits improve it a lot. This guide covers how to keep context intact in long conversations and when to move to a fresh chat.
🟢 Current lineup (September 2026): Claude Opus 5.5 / Claude Sonnet 5 / Claude Haiku 4.5 (higher tier: Claude Fable 5.1). · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
- When context drifts in a long chat, re-inject a short summary of the key decisions to bring focus back.
- Move to a fresh chat when the topic shifts a lot or a wrong assumption keeps following you.
- When summarizing, copy important numbers and decisions verbatim to avoid distortion.
- The context window (how much context the model handles at once) varies by product and version — e.g., Opus 5 is 1M tokens.
Why context fades
A language model considers the whole conversation, but when a chat gets very long, earlier content can carry relatively less weight and key instructions can get buried. The length a model can handle at once (the context window), and how it behaves, varies by model and version, so check the official docs for exact limits.
Four ways to keep context
- Re-inject a summary — briefly restate the important decisions and conditions partway through.
- Restate instructions — repeat must-follow rules (format, tone, constraints) periodically.
- Split topics — start a new chat for a task of a different nature.
- Attach material on the spot — re-paste reference content when needed rather than relying on memory.
When to start a new chat
Move to a new conversation when you see signs like these:
- The topic shifts substantially.
- A wrong assumption from earlier keeps following along — a new chat starts from clean context.
- The chat has grown so long that good answers get harder to obtain.
Summary re-injection example
A quick mid-conversation recap helps Claude refocus:
"To recap so far: (1) the goal is OO, (2) the constraints are OO, (3) we already decided OO. With that in mind, do the following."
Caveats
The amount of context a model can handle at once (the context window) differs by product and version. For example, Claude Opus 4.8 has a 1 million (1M) token context window (per Anthropic's official docs; 200K on Microsoft Foundry). Even with a large window, in very long chats it helps to restate the essentials using the methods above. Also, a careless summary can distort information, so copy important numbers or decisions verbatim.
Frequently asked questions
Why does Claude miss earlier points in a long conversation?
It is a common pattern with language models handling long context. As the context grows, earlier instructions can be reflected less, so restating the essentials helps.
How do I bring the context back?
Re-inject a short summary. For example: "To recap: (1) the goal is OO, (2) the constraints are OO, (3) we already decided OO. With that in mind, continue." This makes it easier for Claude to refocus.
When should I start a fresh chat?
When the topic changes a lot, when a wrong assumption from earlier keeps following the chat, or when the conversation is so long it is hard to get the answer you want. A fresh chat starts from clean context.
How big is the context window?
It varies by product and version. For example, Claude Opus 4.8 has a 1 million (1M) token context window (per Anthropic official docs; 200K on Microsoft Foundry). Even with a large window, restating essentials helps in very long chats.
If you do long tasks often, context management strongly affects output quality. To refine answers you are not happy with, see the iterate-and-refine guide; for more, see the official Anthropic documentation.
To carry context across chats entirely, you can also turn on the Claude memory feature.
Quick Fixes: 4 Ways to Keep Claude on Track (Do These First)
If Claude is ignoring your earlier instructions or forgetting key details, you do not need to start over. Pick the fix that matches your situation below and apply it in your next message.
-
Re-inject a summary at the top of your next message.
Paste a 3–5 line recap of the most important context right before your new request. Claude reads the full conversation each time, but recent content carries more weight — a summary near the top of your latest message keeps the essentials visible.
Success signal: Claude's reply references the details you included without you having to repeat them again. -
Restate your core instruction once every 5–10 exchanges.
If you set a tone, role, or format rule early on (e.g. "always reply in bullet points" or "you are a copy editor"), repeat it briefly when you notice drift. One sentence is enough: "Reminder: keep all responses under 150 words."
Success signal: The format or role snaps back immediately in the next reply. -
Split unrelated topics into separate conversations.
Mixing a coding task and a writing task in one chat forces Claude to balance two contexts at once. Start a new chat for each distinct goal. On claude.ai, select New chat in the left sidebar.
Success signal: Replies feel more focused and stop referencing the wrong task. -
Start a fresh chat when the conversation exceeds roughly 20–30 back-and-forth exchanges.
Very long conversations approach the model's context window limit. (Check the model info page on claude.ai for the current limit — it varies by model and may change.) Before starting fresh, ask Claude: "Summarise the key decisions and open questions from our conversation so far." Copy that summary into the first message of your new chat.
Success signal: Claude in the new chat immediately picks up where you left off with no drift.
Before You Start: What You Need
- Account: Any Claude plan (Free, Pro, or Team) — these techniques work on all tiers.
- Where: claude.ai in a browser, or any Claude-powered app.
- Prior knowledge needed: None. If you can type a message to Claude, you can apply every fix above.
Context window — the maximum amount of text (conversation history + your new message) Claude can read at once. When a conversation fills this window, the oldest messages may be deprioritised. The exact size depends on the model; check the current model card on claude.ai for the latest figure.
Related articles
- Claude Projects, Artifacts & Memory: What They Are and How to Use Them
- How to Write Prompts for Claude: 6 Core Principles from the Official Docs
- How to Use Claude Projects: Knowledge & Instructions to Cut Repetition
- Claude Artifacts: How to Create, Edit, Publish & Share (2026)
- Summarizing Long Documents with Claude: A 5-Step Workflow
- Why Does Claude Miss the Point on Long Documents? 3 Official Long-Context Prompting Techniques
- When Your Claude Character Drifts Mid-Conversation: Keeping Persona Context Stable with Projects