Context window
A context window is the limited amount of text a language model can consider at one time when producing a response. For an AI phone agent, that input can include the system prompt, recent conversation, retrieved reference material, and results returned by tools.
Context is commonly measured in tokens, which are pieces of text rather than exact words. The available window is shared: longer instructions leave less room for call history and retrieved material. When the total input grows beyond the supported limit, the application must remove, summarize, or otherwise manage older content.
A context window is not the same as permanent memory. Information that appeared earlier in a call can influence a response only if it remains available directly or through a reliable summary or state record. Information from a previous call is not automatically present. That distinction matters when buyers hear that an agent can “remember” a conversation.
For phone calls, more context is not always better. A large collection of loosely relevant text can make the important instruction harder to identify and can add processing time. Old tool results may conflict with newer ones. Repeated versions of a policy can produce inconsistent answers. The useful goal is sufficient, current context—not the maximum amount of text the system can accept.
Teams can manage context by keeping prompts focused, storing workflow fields as structured state, summarizing older dialogue, and retrieving only material relevant to the caller's question. Important values should not depend solely on a long transcript. If an appointment date or account identifier controls an action, maintain it explicitly and confirm it when the risk of error is meaningful.
In practice on ThunderPhone, published pricing treats prompt length as a configurable cost driver. Long prompts add 1–2¢ per minute for each additional 10,000 tokens after the first 5,000, and the builder displays the current all-in rate. This makes prompt size both a conversation-design concern and a billing consideration.
When testing a voice agent, run long calls with corrections, topic changes, retrieved answers, and tool responses. Check whether earlier instructions still hold, updated facts replace old ones, and the agent can continue without re-asking settled questions. These tests reveal context-management problems that a short demonstration will not.