How Claude Code Manages Infinite Conversations in a Finite Context Window

Claude Code conversations have no turn limit. You can work for hours — reading files, running tests, debugging, iterating — and the conversation just keeps going. But the model has a fixed context window. At some point, the accumulated messages exceed what the model can process in a single API call. The system needs to compress the conversation without losing critical context. Here's how it works, from the source code. The Problem The naive approach is truncation: drop old messages when the window fills up. This fails immediately. A conversation about building an authentication system might reference a design decision from 50 turns ago. Truncate those turns and the model forgets the decision, re-asks the question, or contradicts what it said earlier. A better approach: summarize. Replace the old messages with a summary that preserves the essential information. But summarization introduces its own problems: What to preserve? File paths, code snippets, user preferences, error resolutions