Context

Context is everything the model can see when generating a response: the system prompt, the conversation history, documents you’ve pasted in, tool definitions and their results — the entire input sequence. It’s the model’s working memory, and it’s the only memory it has. An LLM is stateless between API calls; it doesn’t “remember” your last request unless you send that history again. What isn’t in the context, for the model, does not exist.

This statelessness is the single most common misconception that trips up developers building on LLMs. A “conversation” is an illusion your application maintains by resending prior messages on every call. It also means context is your main lever for controlling behavior: the model learned general capabilities during training, but everything specific to your app — the user’s data, your business rules, retrieved documents — must be placed into context at request time. Deciding what goes in, in what order, and what gets left out is increasingly called context engineering, and it’s a core skill for building agents that stay coherent over long sessions.

Practically, you’ll manage context in code every day: maintaining the messages array you send to the API, injecting retrieved chunks in RAG pipelines, appending tool results in agent loops, and summarizing or pruning old turns as conversations grow. The next topic covers the hard limit on all of this — the context window.