A context window is the amount of information a model can process in one request, usually measured in tokens. It limits what can be available for that response; it is not the same as a chat application's saved history, memory, or document library.

What counts toward the window?

Instructions, included conversation messages, source passages, tool results, and model-specific representations of other media can consume context. Input and generated output share capacity in many model interfaces, and a separate output limit may also apply. Check the limits for the particular model and API.

What happens in a long conversation?

The application decides how to handle a conversation that grows beyond the available window. It may reject a request, truncate content, summarize earlier turns, or retrieve selected information. Older messages are not necessarily deleted from the saved chat, but their original wording may no longer be present in the model's current input.

Anthropic's context-engineering article describes techniques including compaction, structured notes, and retrieval. Those are application strategies, not evidence that every chat product handles long conversations the same way.

Why put critical context early?

Stating the goal and constraints before the source material makes the request easier to interpret. Keep important requirements concise and restate them when changing tasks. Position alone cannot guarantee retention: what matters is which material the application actually includes and how the model uses it.

A workflow for long documents

  1. Define the question and which sections could answer it.
  2. Supply focused passages with section or page labels.
  3. Ask for a claim-to-source table and mark unsupported claims.
  4. Summarize each section before combining results, retaining source labels.
  5. Compare the final answer with the original passages, especially exceptions and numbers.

Use the research-summary workflow to track omitted material and distinguish a model's active window from general background context.

Common mistakes

  • Measuring the window in words or characters rather than the model's token accounting.
  • Assuming that a larger window guarantees better reasoning or complete recall.
  • Equating an uploaded file with every part of that file being used in every response.
  • Relying on “remember this” without checking the next answer against the original requirement.

Sources and further reading