How Claude thinks
Context windows, memory, and what understanding really means
The one fact that explains everything
Claude doesn’t think continuously. Every time you send a message, Claude reads the entire conversation from scratch — top to bottom — and generates a response. There is no background process, no persistent memory between messages. Each response is a fresh read of everything visible in the conversation.
This single fact explains almost every Claude behavior you will encounter.
The context window
The context window is Claude’s working memory — everything Claude can see at once. It has a size limit measured in tokens (roughly 3/4 of a word each). Everything in the window influences Claude’s response: your messages, Claude’s replies, documents you’ve pasted, system instructions, and tool outputs.
Key principle: What’s in the window = what Claude knows. If you mentioned your company name three conversations ago, Claude has no idea — that conversation isn’t in the window.
What Claude does and doesn’t remember
| Claude DOES remember | Claude does NOT remember |
|---|---|
| Everything earlier in this conversation | Anything from a previous conversation |
| Documents you pasted this session | Files you uploaded in another chat |
| Instructions you gave earlier in this chat | Your preferences from last week |
| Tool results returned this session | Who you are, by default |
Memory features work by injecting stored facts into the context window at the start of each session. Claude isn’t “remembering” — it’s reading a file that was pre-loaded before you typed anything.
What understanding actually means
Claude doesn’t understand language the way humans do. What it does is closer to sophisticated pattern completion trained on an enormous amount of human writing. When you write a prompt, Claude is asking: given everything in this window, what is the most coherent, helpful continuation?
- Clarity compounds. The more precise your prompt, the better the pattern Claude is completing toward.
- Framing matters. “Write a summary” vs “Write a summary for a non-technical executive” pulls from very different training patterns.
- Claude can be confidently wrong. Always verify specific facts, URLs, and dates.
Practical takeaways
- Put important context early. Front-loaded context sets the stage for everything that follows.
- Long conversations can drift. Early instructions get diluted by later content. A fresh conversation with a well-written opening often beats a 40-message thread.
- You can correct anytime. Claude re-reads everything including your correction.
- Pasting documents = brain transplant on that topic. Drop in a 20-page brief and responses will be grounded in it for the rest of the session.
Try it: Start a fresh conversation. Give Claude a client name and project budget in your first message, then after several exchanges ask “what’s the budget we’re working with?” Watch it pull from earlier context.