Feature request for product/service
AI Models
Describe the request
Hi everyone,
Right now, in a prompt session, when the context window is close to full, Cursor automatically summarizes the context. There is also a /summarize command that allows us to trigger this manually when needed. Thanks to a recent update that introduced the “context usage breakdown” feature, I’ve had a chance to better understand how Cursor’s Context Summarization actually works. It compresses the “Conversation” context from ~190k tokens down to just ~1k.
Don’t get me wrong - I am completely fine with context summarization. However, I want to have more control over how Cursor summarizes that context. When running a workflow - for instance, implementing the code for a somewhat complex feature - I often hit the context summary limit a few times. The problem is that after the context is summarized, the AI model’s performance seems to degrade significantly, not in the coding ability, but in the requirement understanding and skill instruction following. Sometimes it manages to push through and correctly follows the initial instructions; other times, it completely loses track of what it was doing, goes rogue, and starts messing up the codebase. In worst-case scenarios, it just stops halfway through a task and says, “I’m done, please verify,” when the job isn’t even 50% finished.
It seems like Cursor’s approach to summarizing context is on the right track, but is it just relying on the model’s own reasoning to decide how to summarize? This would explain why the quality of the summary is so inconsistent across different runs.
Could the Cursor team provide us with a mechanism to specify or override how context is summarized? For example, a way to define:
- What must be kept as-is (e.g., system/skill instructions, mandatory files to read…).
- What can be concisely summarized (e.g., files already modified, steps completed, active bug fixes…).
- What can be completely purged from the context (e.g., read skills that are no longer needed, files that were opened but are irrelevant to the current task…).
This could be implemented as a skill, a hook, or an editable system prompt. That would be amazing!
We don’t need the context aggressively compressed down to ~1k tokens if it means the Agent essentially gets amnesia afterward. We need a “selective context summary mechanism” - even if it only compresses down to ~10k tokens - so the Agent actually stays as smart as it was at the start of the session.
Operating System (if it applies)
Windows 10/11
MacOS
Linux