Describe the Bug
I’ve been complaining about this all week: the new conversation summarizer is painful to use. It constantly causes the AI to drift, lose its working documents, lose its reference materials, and just straight-up waste tokens. Now I have to re-prime the AI every single time, drag the documents back in, re-explain where we were, and copy-paste the old chat just to restore even a basic level of understanding.
When I ask the AI, “What do you think we were doing based on your summarization?”, nine times out of ten it answers with something like, “Oh, we were doing this, this, and this,”—which is either just the first and last things we discussed (completely ignoring everything in between) or it’s completely off-base and has no idea what we’re talking about. All nuance is stripped away, important information is lost, and it ends up repeating the same scripts over and over with only minor name changes.
It feels like talking to someone with short-term memory loss who jots down random words with no actual context—leaving them incapable of maintaining a real conversation.
If this compressor is absolutely necessary (which makes no sense to me since I’m paying for the tokens and should decide how to use them), then give us control over what it compresses, where it compresses, and how aggressively it compresses. For example:
A value-based compressor that judges the importance of statements and decides the compression level with the ability for the agent to expand the previous conversation based on a confidence value.
A semantic linking map where summaries are tied directly to the actual conversation, with an agent capable of expanding them on demand
Possible user settings could include:
Only compress the oldest half of the conversation
Only compress large messages
Only compress a specific percentage of the context window (top, middle, bottom — and by how much)
Just give us any level of control if you’re not going to let us turn this useless thing off. This is exhausting. I’ve spent way too much time and energy fighting the AI’s drift, hallucinations, and the mountains of technical debt it’s creating. Honestly, if there’s no toggle or compression controls in the next couple of weeks, I’ll be canceling my team’s plans. This is making us have to back track constatly and redo the simple things manually that we did not have to worry about.
Steps to Reproduce
Have a long conversation of developing something, then leave the ai to run by itself working with the documents created based on that as well as the context.
Expected Behavior
The ai to remember anything of value, not drift and generate mountains of technical debt. Not waste way more tokens re-priming the ai over and over, loading documents back in, and keeping the actual objectives, too do lists and more relevent and contextually understood?
Operating System
Windows 10/11
Current Cursor Version (Menu → About Cursor → Copy)
Version: 1.3.7 (user setup)
VSCode Version: 1.99.3
Commit: 5d56108ff2401c28f239aa2a527ef424ce7719a0
Date: 2025-07-31T04:16:31.276Z
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Windows_NT x64 10.0.26100
Additional Information
You could even grab the most important conversational pieces and attach them directly to the too do list items so there is a genuine documented link to the task it is attempting to do.
Does this stop you from using Cursor
Yes - Cursor is unusable