Your context management is seriously lacking. I open new windows all day long; after only a few rounds of conversation, the current session becomes unusable and I have to start over in a new window. This makes the product very frustrating to use day to day.
Steps to Reproduce
Steps to Reproduce
Open Cursor on a real project (large repo with many files, e.g. monorepo / multi-package workspace).
Start a new Agent or Chat session in one window.
Work through several rounds of back-and-forth (e.g. 5–10 turns): ask questions, request code changes, let the agent read/edit multiple files, run terminal commands, and continue in the same thread.
Keep extending the same conversation instead of starting a new chat.
Observe that after a few rounds:
responses become slower or less accurate,
the agent “forgets” earlier context or repeats work,
tool calls / file edits miss details from earlier in the thread,
or the session feels unusable and you must open a new window / new chat to continue productively.
Repeat daily on normal development work — the pattern recurs: every few rounds of work, a fresh window is required.
Expected behavior: One session should retain useful context across a normal work session without forcing frequent new windows.
Actual behavior: Context degrades quickly; I routinely open new windows/chats after only a few rounds of work.
Hey, thanks for the detailed report. I want to separate two things right away, because this mixes expected behavior and a possible bug.
What you described after 5 to 10 turns in one thread, the agent starts forgetting earlier details, repeating work, and responding more slowly is mostly about the limited context window. Any LLM has a limited window, and in long conversations older turns get compressed via summarization, so some early details get lost. Starting a new chat for a new task isn’t a workaround, it’s the recommended way to work.
What actually helps keep quality steady throughout the day:
Start a new chat for each logically separate task so the context stays dense and relevant
Keep a thread focused, don’t mix unrelated tasks in the same conversation
Explicitly point to the right files via @ references instead of having the agent re-scan for context every time
Use Rules Rules | Cursor Docs for repeated project instructions, they don’t depend on thread length
Use Plan mode for larger tasks to lock in a plan before making edits
Now, if what you’re seeing isn’t general degradation but a specific failure, like the agent ignoring context in a fresh short chat, or clearly answering the previous question instead of the current one, that’s different and I’d like details. Please send the Request ID for that case right upper corner of the chat > Copy Request ID and a short note on what the agent missed. Then I can look into it properly.
I’ve restarted windows countless times, but it’s gotten so bad that I now need a new window just to finish a single feature. This happens with both GPT 5.5 and Composer 2.5 — especially Composer 2.5. After only three rounds of conversation, it becomes unusable and keeps grap-ing files. What’s going on? It’s worse than before! requestid：305ff8ed-09af-47a1-90b0-0a7b0ae41b11