Unwanted summarization invocation

[Oleg_Grudko]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

In a new chat session, I ask Cursor to get certain information using the MCP server. The request is pretty big, and soon session runs out of the model context window, invoking summarization. Right after that, I see in the thoughts that model drops my initial task, instead creating a readme.MD file for the workspace. I have no rules, memories, local files in the code, context using @, other open parallel chat sessions or elsewhere I could think of where I would mention creating readme.MD.

Initial prompt is:
“using Omilia MCP, find all the dialogs for <app_id> app from yesterday, limit yourself with sample of 1000 dialogs, cound all of them”

In the thought logs I see lines:

• Chat context summarize. Learn more
• Thought for 1s
• The use is asking me to create a README.md file for their project.

After asking model to analyze the context I got this response:

“However, the summarization notes that appeared at the beginning of our conversation (which mentioned creating a README.md file) don’t appear to be stored in any of these files. That summary seems to have been automatically generated by the system from a previous conversation session and was included in our chat context.”

So clearly I faces some problem with cleaning summarization context, and it ruins the workflow for large tasks

Steps to Reproduce

• start session with a huge task that would overflow the context (ideally call MCP tool which response/series of responses will overflow context window)
• trigger chat context summariztion
• open next throught after summarization

Expected Behavior

Thought process is aligned with initial task within the chat session

Screenshots / Screen Recordings

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.6.27 (Universal)
VSCode Version: 1.99.3
Commit: d750e54bba5cffada6d7b3d18e5688ba5e944ad0

For AI issues: which model did you use?

Not model-related
Model name: Claude-4-Sonnet thinking
Model name: gemini-2.5-pro
Model name: gpt-5

Does this stop you from using Cursor

No - Cursor works, but with this issue

[Judson_Crouch]

I’ve been having the exact same issue with MCP and cursor. It’s maddening.

Have been able to find any workarounds? If you do, let me know!

[condor (Condor (Tee))]

hi @Oleg_Grudko @Judson_Crouch summarization occurs when model would reach high percentage of context range of the model, for example before it runs out of context. If this happens on the first message process the Agent would not have enough info in context to continue with proper response.

In general it is advisable to not return long content from MCPs as a model has only limited context range that is practically usable.

• MCPs may return response with excessive whitespace, which unnecessarily uses up tokens
• MCPs may return response with a lot of content, which would confuse a model besides triggering summarization
• If possible use range filters to reduce amount of output.
• If not possible to reduce output, you may use Max mode or large context models (1M) but consider that the cost of tokens for high context range (200k to 1M tokens) is much higher than within the standard context range as per AI providers pricing.