As a chat fills up, older details get summarized away and quality drops; managing it keeps the agent sharp.
Everything in a chat (your messages, files the agent read, command output, rules, tool descriptions) takes space in the model's context window. Cursor shows this with the context ring in the chat input. Click it for a breakdown by category: system prompt, tools, rules, skills, MCP, subagents, summarized conversation and conversation.
When the window nears full, Cursor summarizes older content to make room. Summaries keep the gist but can drop exact details, such as a constraint you mentioned an hour ago.
Habits that help:
- One task per chat. Start a new chat when you switch to unrelated work. Long chats about many things carry stale context into every request.
- Carry context deliberately. In a new chat,
@-mention the earlier chat (Chats) or the plan file instead of continuing a bloated conversation. - Side questions: a side chat (
/sideor/btw) answers a quick question without adding it to the main thread. - Don't over-attach. Mention the files that matter, not whole folders "just in case".
- Restart after repeated failures. If you corrected the agent two or three times on the same thing, the chat is full of failed attempts. A fresh chat with a sharper prompt that includes what you learned usually wins.
- Put lasting knowledge in rules, not in chat history, so every new chat starts with it.
Context also affects cost: every request resends what is in the window, so a huge chat makes each message more expensive as well as less accurate.
Remember
- The context ring shows how full the window is and what fills it
- Older content gets summarized near the limit; details can be lost
- New chat per task; reference old chats or plans with @; durable facts go in rules
Watch out for
- Running one chat all day across unrelated tasks, so every request drags along irrelevant history.