Glossary · Core Concepts

Context Window

Core Concepts 新手

30-Second Version · For the impatient

The maximum amount of text Claude can hold in memory during a single session — everything beyond it gets forgotten.

Full Explanation +

01 · What is this?

The Context Window is the total amount of text Claude can see in a single conversation — every message you send and every reply Claude gives consumes this fixed space. Think of it as a fixed-size whiteboard: once full, the oldest content on the left gets erased to make room for new content on the right. Claude doesn't selectively forget — it physically cannot see content that has scrolled out of range. If you gave detailed instructions at the start of a conversation, once those instructions fall outside the Context Window, Claude behaves as if it never saw them — and critically, it doesn't know it doesn't know, which is the most dangerous aspect. Different Claude plans have different Context Window sizes, but the core problem is the same: the further you go into a conversation, the less of the beginning Claude can reference. This is why many users feel Claude gets dumber over time — it's not getting dumber, it just can't see what you said earlier.

02 · Why does it exist?

Without a Context Window mechanism, a language model would have complete amnesia on every reply — it couldn't understand what the plan we just discussed referred to. The Context Window is the fundamental mechanism that makes multi-turn conversations possible. But it also creates a cost structure problem: every API call requires the entire conversation history to be re-sent as input tokens. The longer the conversation, the higher the cost per reply — you're not just paying for this one response, you're paying for reading through the entire conversation history again. This is why Context Window management is a core architectural decision when building AI applications: you need to find the balance between conversational coherence and API costs.

03 · How does it affect your decisions?

The Context Window directly affects three of your core decisions: First, put your most important instructions last. Forgetting starts from the earliest content — what you put last gets pushed out last. Put your most critical constraints in recent messages, or use System Prompt to make them persist. Second, use RAG for long documents instead of pasting directly. A 50-page contract pasted into a conversation consumes massive Context Window space. Using a RAG system to store documents in a vector database and retrieve only the most relevant chunks is far more efficient. Third, start new conversations regularly. Don't let one conversation grow indefinitely. Note important conclusions and instructions, then paste them into a new conversation to reset the Context Window.

04 · What should you do?

Three things you can do immediately: Use Claude Projects. Projects let you permanently store background documents and instructions in a project, without consuming the conversation's Context Window. Every new conversation automatically carries this background. This is the lowest-barrier solution for individual users. Monitor conversation length. If you use the API, track the input token count per call. When input tokens exceed 70% of the Context Window, it's time to start a new conversation or compress the history. Build a conversation summarization habit. For tasks requiring long-term continuity, periodically ask Claude to summarize current discussion conclusions, save the summary, and paste it as background when starting a new conversation.

Real-World Example +

Scenario: You're using Claude to analyze a 60-page legal contract. Wrong approach: Paste the entire contract into the conversation, then ask 20 questions. By question 15, you ask about the penalty clause in Article 3. Claude gives a vague answer without referencing the contract text. You think Claude is making things up, but actually the contract content has been pushed out of the Context Window — Claude can no longer see Article 3. This mistake is insidious: Claude won't say it can't see the contract anymore. It continues answering, but answers start becoming inaccurate. Correct approach 1 (RAG): Store the contract in a vector database. Every time you ask a question, the system retrieves the most relevant contract sections. Claude always sees the most relevant content, unconstrained by the Context Window. Correct approach 2 (Claude Projects): Upload the contract to Claude Projects with analysis instructions. Every subsequent conversation has access to the contract. Correct approach 3 (batch processing): Ask the most important 5 questions first, start a new conversation with the contract summary and next batch of questions.

Diagram

Feel free to share. Please credit the source.

Common Misconceptions +

✕ Misconception 1

× Myth 1: Bigger Context Window = better, always prefer the largest available. A larger Context Window means higher API costs per call and slower response times. It also encourages users to stuff everything in, leading to runaway costs. The right approach is choosing the appropriate size for your task needs, and supplementing memory with Projects and RAG.

✕ Misconception 2

× Myth 2: Claude forgetting things is a bug — it should remember all conversations. This is an architectural property of language models, not a bug. Stateless design makes each API call independent. True permanent memory requires external storage (Projects, databases, vector stores) — not a bigger Context Window.

The Missing Link +

Direct Impact

Larger Context Window = more coherent conversations + higher API costs + slower responses. Smaller Context Window + external memory (RAG / Projects) = lower cost per call + faster responses + requires additional architectural work. For individual users, Claude Projects is the lowest-barrier solution. For developers, RAG architecture is the long-term scalable choice.

Ask a Question

Related Terms

Useful Resources

Claude API Status → Model Pricing → Prompt Playground → Token Counter → MCP Servers → LLM Benchmarks → Model Comparison →