The context window is the total text Claude can see: your messages, replies, file contents, and tool outputs. It has a fixed size. When it fills, Claude starts losing earlier parts of the conversation.
You'll see the signs before it hits: Claude forgets earlier decisions, repeats answered questions, or contradicts what you agreed on. Large file reads and verbose tool output burn context fastest. I've seen a full git diff on a large PR chew through as much context as an entire hour of conversation.
MCP servers add hidden overhead. One developer documented tokens consumed by servers before typing a single message. Run /insights to see a token breakdown by category and /cost to see exact token counts.