Context keeps summarizing at 10-20% of total context window

[context_is_broken]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

ultra user here: every time I try loading any model, give it a prompt, and have it execute on the prompt, agent is only able to get past ~10% of context window before summarizing context (in the case of using GPT 5.5, context usage explodes to 50% of total window despite not having analyzed any bulky docs, and this burns API tokens), seemingly forgetting what it just read, and then it gets stuck in this loop. have tried with several different frontier models + composer 2.5 + plan/agent/debug/multi-task mode. is this a widespread issue or any experience with this issue?

Steps to Reproduce

on any new conversation, this bug will arise once roughly ~20% context has been reached.

Expected Behavior

We want to be able to reach full context window usage + not catastrophically forgetting prior context after “summarizing chat context” fires + no context explosions with frontier models.

Screenshots / Screen Recordings

Operating System

Windows 10/11

Version Information

Version: 3.8.11 (user setup)
VS Code Extension API: 1.105.1
Commit: e56ad3440df06d22ca7501e65fd518e905486ef0
Date: 2026-06-18T01:40:18.333Z
Layout: editor
Build Type: Stable
Release Track: Default
Electron: 40.10.3
Chromium: 144.0.7559.236
Node.js: 24.15.0
V8: 14.4.258.32-electron.0
xterm.js: 6.1.0-beta.256
OS: Windows_NT x64 10.0.26200

For AI issues: which model did you use?

Composer 2.5, GPT 5.5 1M Extra High (keeps exploding from 5% context usage to 50% usage within seconds without having analyzed any documents), Opus 4.8 1M Max

Does this stop you from using Cursor

Yes - Cursor is unusable

[kevinn (Kevin Neilson)]

Thanks for the detailed report, and sorry for the frustration. Two things are combining here.

1. The window is smaller than 1M unless Max Mode is on**.** The 1M figure only applies when you select Max Mode plus the 1M context variant in the model picker. Otherwise frontier models run on their standard window (200k-300k tokens), and Composer 2.5 runs on 200k token context window. Summarization fires near the top of whatever window is active, so it’s not possible to use, for example, 199k tokens in practice on Composer 2.5 because self-summarization will fire before that.

2. The bigger driver: searches are pulling in huge amounts of output. Looking at your recent sessions, the context is filling from tool output, not documents. Several searches walked environment and library folders (for example conda/miniforge environments and Python virtualenvs) and a single search returned over a million characters. One result like that fills a large chunk of the window in one step, which is why usage jumps to 50% “within seconds” and the agent then summarizes and seems to forget.

The change that helps most is keeping those folders out of search. Add a .cursorignore at your project root, for example:

**/miniforge/**
**/miniconda/**
**/.venv*/**
**/venv/**
**/node_modules/**
**/build/**
**/*.so

You can also have the UI build your .cursorignore for you by going to Cursor Settings → Indexing and Docs and then Editing the list here:

Try the .cursorignore first, since it’s almost certainly the largest contributor.