By Teresa Torres · producttalk.org · @ttorres on X · LinkedIn
Teresa Torres explains "context rot," the pattern where large language model performance degrades as the context window fills up with conversation history and input. Drawing on recent research (Liu et al. 2023, Paulsen 2025, Veseli et al. 2025, Du et al. 2025), she describes how models lose tokens in the middle when less than half full and lose the earliest tokens when more than half full. The piece argues that managing the context window is a third lever alongside prompt engineering and context engineering. It notes that web chat interfaces hide context usage, leaving starting a fresh chat as the main fix, and introduces Claude Code as a tool offering far more visibility and control. The article is cut off at a paywall before its practical tactics are fully laid out.
01Key takeaways
- Treat the context window as a third lever, alongside clear prompting and providing the right context.
- Long conversations degrade model performance, so long chats can lead to ignored instructions and false success claims.
- Start a fresh chat when you switch topics, when the model misbehaves, or when a conversation passes roughly 15 messages.
- Ask the model to summarize a long conversation and carry that summary into a new chat.
- Use tools that show context usage so you can match how full the window is to the complexity of the task.
02Key sections
- The phenomenon
- The author describes noticing that AI agents repeatedly claimed fixes that never happened, and that starting a new conversation often resolved the problem. She frames this as a broad issue across tools.
- What a context window is
- The context window is the model's short-term memory, with a fixed size that varies by model and includes system prompts and the full conversation history on every turn.
- Research on context rot
- Successive studies show performance degrades as input grows, with position effects shifting depending on how full the window is, and a 2025 experiment suggesting the problem is input length rather than retrieval.
- Managing context in the browser
- Web chat tools hide how full the context is, so the main practical lever is starting fresh chats when topics change, outputs go wrong, or conversations grow long.
- Visibility in Claude Code
- Claude Code displays context usage and offers commands to inspect, clear, and compact context, letting users calibrate how full the window should be for different tasks.
03From the post
“Have you ever noticed that AI gets worse the longer you talk to it? I first noticed this when I was trying to fix bugs on Replit. I'd spin in an endless cycle of asking the agent to fix something and it would report back that it fixed it,”
04Frameworks mentioned
Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.