That is hard to judge as I do tend to put for that specific project like 5 files into the context because I can (I assume), but not more like 400KB in filesize total.
Can’t recall, whether the “create new chat” hint was already showing.
And my rules are usually more the opposite of what was happening: stay on track, do not overcomplicate things, not “add more QoL stuff while you’re at it”. 
Good question, but I can’t give any good answer to it. 
1 Like
Version: 0.48.7 (user setup)
VSCode Version: 1.96.2
Commit: 1d623c4cc1d3bb6e0fe4f1d5434b47b958b05870
Date: 2025-04-02T05:04:28.547Z
Electron: 34.3.4
Chromium: 132.0.6834.210
Node.js: 20.18.3
V8: 13.2.152.41-electron.0
OS: Windows_NT x64 10.0.26100
those leaking chats happen more often then you think, also readed i’m not the only one when cmd get’s stucks, also today especialy there is a lot of edit_file errors, you get sometimes bypass those edit tool errors with new chat, also claude ai today especialy dosen’t work at all its just loading and loading and loading, to bypass this its good that i have my own claue api key so i switch one message with api one again with cursors agent and it come back sometimes to normal, somwhere of cursors teams member writted in here in forum that that error edit tool is happening in long chat’s no it can hapen almost instantly in new chats also, (Error calling tool ‘edit_file’.), nowadays, i don’t know your servers cannot take the pressure or what is happening but the ai you can say its unusable, waiting for the updates hopfully it will fix it but don’t want to downgrade also, also if possible extent how much ai remembers in chat history not like now 3 messages and he starts again to analyze my whole project to understand what is happening, good idea would be implement with persons decision that it would use local space for ai to hold all data and what he did, i saw this i think in feature requests it was very detailed written.
When I first started using Cursor I had most all these problems.
Then I added documentation files.
- From within Cursor: File / New Text File
- Write out an overview of your project.
- Break down your project into screens.
- Break down the screens into “User Stories” eg., “User logs into the system”
- Break down the user Stories into components that are needed to accomplish the task.
Save the document as a “.md” file type in your project root folder.
You should see fewer haulucinations.
I will be launching “In Focus” that will handle this from a series of “Segments” guided by AI. If you’re interested DM me.
what do you mean with “Write out” ? does it requires a keyboard? the purpose of Cursor is the opposite
1 Like
(wink)
You’re really into the true “Vibe Coding” aren’t you.
Stare at the screen and AI reads your mind. Let me know how that works for ya.
2 Likes
Hey @danperks ,
I’ve got a current example for this thread’s issue and remembered to get a request ID for you:
b5e89e8e-d3d9-487a-a8f3-5509cb7fa28a
(Cursor’s Gemini Pro 2.5, but not max).
It was a rather simple “search and replace” task across files (literals to enums), where suddenly a test script got 200+ lines of additional (failing) tests added.
I then asked Gemini to explain what happened, and its little judgemental reply made me laugh a bit, tbh: 
Quoting from its “thoughts”:
"The user is questioning the large number of added lines in server.test.ts during the first edit_file call for that file. I claimed that only the error code replacements and the import addition were intended.
…
** The instructions argument was: “Add ErrorCode import and replace hardcoded error numbers with ErrorCode enum members.”*
The less intelligent model applying the edit_file likely interpreted the presence of the full test cases in the code_edit as instructions to add those test cases, especially since they weren’t marked with // … existing code … The model saw the ErrorCode change within those blocks and assumed the entire block was part of the change I intended."
Maybe this helps improve the system if the team can view the convo or DM me for the full chat. 