Context Window
A LLM's working memory: how much text it can 'see' at once (conversation + instructions + documents). The larger the window, the more context you can share. Past ~40% fill, quality collapses (dump zone): the LLM retains mostly the beginning and the end of the window, the middle blurs and hallucinations rise ('lost in the middle' pattern). A 1M-token window does not mean 1M usable tokens, the practical ceiling is typically around 400k.
Strengths
- Claude Sonnet 3.7: 200k tokens, enough for a medium codebase
- Operational heuristic: aim for <40% fill before compaction, past that, hallucinations rise
Limitations
- Large windows cost more in tokens, check pricing
Best for
- Understanding why an agent loses track on long sessions