Learn
What is a context window?
A context window is the maximum amount of text (tokens) a model can consider in one request — prompt, retrieved documents, tool results, and output combined. It bounds how much an agent can 'see' at once; long-horizon work requires memory outside the window.
In depth
Context is the model's working desk. Everything on the desk must fit: system instructions, user message, RAG chunks, prior tool outputs, and the answer being generated. When the desk overflows, older content is truncated or summarized — silently, unless you design around it.
In 2026, frontier models offer 128k–1M+ token windows, but usable context is smaller: quality degrades in the middle, cost scales with length, and latency grows. Production agents treat the window as precious: retrieve only what matters, compress history, and store durable state in git, databases, or skill files.
Context engineering — what you put in the window and in what order — often beats raw window size. Put constraints and examples early; put bulky reference material in retrieval; keep tool outputs structured and short.
Neuro OS roles use scoped memory layers and external artifacts so work continues across sessions instead of stuffing one giant prompt.
Examples
- 200k window — full competitor report in one pass for analysis
- Truncation — drop oldest chat turns when limit approached
- Summarize-then-continue — compress prior sprint into bullet memory
- RAG — only top-5 chunks enter context, not whole corpus
- Neuro OS skill file — procedure lives in git, not in every prompt
Related terms
FAQ
Is a bigger window always better?
No. Larger windows cost more, run slower, and models still lose focus. External memory + retrieval usually scales better.
How do agents handle work longer than the window?
Heartbeats, file artifacts, databases, and summarization — see long-horizon agents.
Run governed agent roles on a company OS — not only definitions in a glossary.