Claude Code may offer a 1 million-token context window, but Anthropic says long sessions can still degrade over time. In an official blog post, Anthropic technical team member Thariq Shihipar described the issue as “Context Rot” — a state where older material lingers in the prompt, attention gets diluted, and the model starts missing requests, answering less precisely, or even undoing earlier fixes.
A larger context window does not remove context decay
The context window is the total amount of information Claude can consider at once, including the chat history, files it has read, and tool calls made during the session. Claude Code’s window is described as roughly equivalent to 7.5 million English characters. That scale is large, but the post argues that size alone does not solve the problem. As more stale logs, failed debugging attempts, and outdated instructions remain in context, they can interfere with the task currently in front of the model.
Anthropic said that once the context gets close to its limit, the system automatically triggers compaction, summarizing earlier conversation history to free space. The problem is timing. According to the post, compaction often happens when Context Rot has already weakened model performance, raising the chance that important details get dropped during summarization.
Five official options for handling long Claude Code sessions
The first option is Continue, which simply keeps working in the same session when the task is still on track and the context remains useful. The second is Rewind, available by double-tapping Esc or entering /rewind, allowing the user to jump back to an earlier message and restart from that point. The idea is to keep useful reading and discoveries while removing the record of failed attempts.
The third tool is Clear, triggered with /clear, which drops the entire context and starts fresh. It requires the user to decide what information should be carried into the next session. The fourth is Compact, using /compact to summarize prior history inside the same session. Anthropic said users can guide that summary with instructions, such as keeping an auth refactor and excluding test debugging details.
The fifth option is Subagents. A subagent handles a subtask in a clean context window and returns only the output that matters, while the intermediate noise stays isolated from the main session. Anthropic’s examples include reviewing another codebase’s authentication flow and summarizing it, or validating test results for a specific change.
The rule of thumb: new task, new session
Anthropic’s decision guide is straightforward. Continue when the task has not drifted. Use Rewind when Claude has gone down the wrong path. Choose Compact with instructions when the session is getting crowded. Use Clear for a completely new task. If only the final answer is needed, assign the work to a subagent.
Shihipar’s core recommendation is simple: a new task should start in a new session. Claude Code can hold a very large amount of context, but related work and unrelated work should not be treated the same. For developers, active session management can preserve output quality better than waiting for automatic compaction to step in.