CursorContext mistakesLevel 2: Context

When Should You Start a New Chat in Cursor?

Start a new chat when you switch tasks, when you finish a feature, or when Agent keeps going in circles, even if you're mid-feature. Keep the current one only while it's on the same work and its earlier context is still useful. To bring the history along, mention the old chat with @Chats.

Prompt to copy
Continue [the work] from @Chats [pick the previous conversation].

Current state: [what's done, what's committed]
What went wrong before: [the approach that failed, so you don't repeat it]
Next step: [one specific task]

Read the current code rather than relying on the old conversation where they disagree.
Keep goingsame chat
  • Same feature
  • Useful earlier context
  • Agent is on track
Start freshnew chat
  • New task
  • Agent making mistakes
  • Going in circles
Cursor's course gives this guidance on continuing versus starting fresh.

Try it now

  1. 1Press Cmd+N (Ctrl+N) to start a new chat when your current task is done.
  2. 2Type @Chats and pick the old conversation if you need its background, then send the prompt above.
  3. 3Check how the agent starts. It worked if it continues from your stated next step, without retrying the approach that failed.

Use it when

When the task changes, after two or three failed attempts at the same fix, when the agent contradicts an instruction it followed before, or when the context ring shows it's nearly full.

Skip it when

Don't restart in the middle of a working feature just because a chat feels long. Cursor's course says to continue while the agent has useful context. Restarting would also throw away what it has read, and it would have to find it all again.

Why it works

Context is the agent's working memory, and everything said in a chat stays in it, including wrong turns. Cursor's course compares this to trying to remember what someone said three hours ago. When the window nears its limit, Cursor compresses older messages into a summary, so details may get lost. Old mistakes also keep influencing what comes next, and a stale approach may seem settled. A fresh chat removes that, while project rules still load at the start.

Example

Cursor's course shows a follow-up along these lines: reference the old auth refactor chat, say the review feedback on token expiry is dealt with, and ask for the next piece, refresh token rotation. The new chat has the facts it needs, and the back-and-forth about the review isn't taking up space.

Going further

You've got softer options before a full restart. Run /summarize to compress the chat on demand, fork from an earlier message to drop a bad branch, or use a side chat so a tangent never enters the thread. Note that restoring a checkpoint reverts files only, so the failed attempt remains in the conversation. That's why a restart can help even after you've rolled the code back. Long-running agents can auto-summarize, which Cursor's help mentions for pinned conversations, but it isn't a reason to mix unrelated tasks in one thread.

Common mistakes

  • Keeping one chat open all day across unrelated tasks.
  • Repeating the same request to an agent going in circles, instead of restarting with what you've learned.
  • Starting fresh with no summary, so the agent repeats the same failed approach.
  • Believing a checkpoint restore clears the history. It reverts files only.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course