Summarization not working

Support Bug Reports
, ,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

I summarize, then right after the context window is almost full again, without having interacted with the AI or is empty and then full again right after the first interaction. This is a bug I have observed at least since the beginning of this year.

Steps to Reproduce

Summarize an almost full context on Opus 4.6

Expected Behavior

Context window is small after the summary.

Screenshots / Screen Recordings

Chat Summary2.jpg
Chat Summary2.jpg1073Ă—1332 105 KB

Operating System

Windows 10/11

Version Information

IDE: 3.0.13

Does this stop you from using Cursor

No - Cursor works, but with this issue

7

I found a much better solution than to summarize.

Referencing.

Open a new chat and reference the chat you want to summarize and continue with the conversation.

9

Hi @Hans_Muster Thanks, for returning and sharing your experience here - that is very helpful. We’re working to improve the summarization, but glad to hear that opening a new chat with a fresh context window and referencing the prior chat worked out well - that’s a great option and thanks for returning to share your experience.

10

This is not a solution. The /summarize feature is still broken. I don’t want to be creating new chats from scratch 5-10 times on a long conversation, trying to refer the correct previous conversation, and hope it does not pick an older chat in this sequence.

Currently, I am having to workaround this by manually renaming each new chat and adding a serial number to it to keep track of the conversation, and is super painful. This makes cursor super hard to use for anything non-trivial.

I am not asking for an improvement to summarization here. But the /summarize feature appears to be a no-op with no change in context at all. Just that the UI does not show how full the context window is right after you do a /summarize, and it starts showing the original fullness + new contents from the new message right after. Hopefully it’s a silly bug that’s easy to fix.

11

Hey @Govind_Balaji_S

Here is another idea that might help solve the problem for you:

Tell the agent to summarize it into a new file inside the project, then work from that in a new chat.

Repeat that process with that one file once new chats get too big again

This way you will always have one source of truth regardless of how many new chats you open.

Let me know if this helps.

12

Thanks Hans, appreciate the workaround. But I really would like to be able to avoid the overhead of opening multiple “sequential” chats/agents (the list of chat/agents gets messy for me, especially if I fork the chat too from some of these “sequential” chats). I decided to bump this thread, since it looked like Cursor team might have thought this was an issue with how well /summarize really summarizes (a hard problem) vs /summarize being a no-op (hopefully a silly bug that’s much easier to fix).

14

Hi @Govind_Balaji_S and Hans I appreciate you guys following up here.

@Govind_Balaji_S What would really help us investigate this further is a link to a request ID where the summarization isn’t working, but I think your org probably enforces privacy mode. I do see we have one other report of this on Agents view. Are you using agents view or the regular IDE when this summarization does not work as expected?