Context windows and memory myths

What counts toward context

The Claude docs state that everything in the request counts toward the context window: the system prompt, messages, tool results, images, documents, tool definitions, output, and extended thinking. This is why a workflow can feel like it has plenty of room until it suddenly degrades or truncates.

The model overview lists current context and output limits by model. Several current models support 1M-token context windows, while other models are smaller. But model selection is only one part of the problem. The content you include must still be relevant, well-labeled, and not polluted by stale instructions.

Three common memory myths

A long conversation can preserve irrelevant history alongside crucial facts. That can distract the model, increase cost, and bury the active instruction. In chat products, the surface may also manage old context differently from the API. In the API, a developer has to decide what to keep, summarize, or remove.

• Myth 1: If it appeared earlier in the chat, Claude will use it correctly forever.
• Myth 2: A larger context window means the answer is automatically more accurate.
• Myth 3: A context-limit failure is the same thing as a safety refusal.

Context practice that holds up

Keep a durable facts section. Reattach primary sources when they matter. Remove obsolete instructions. Use token counting before sending large jobs. Split tasks when the output itself is large. Ask Claude to produce a short evidence table before writing long prose.

Most importantly, treat model_context_window_exceeded, truncation, and refusal as separate states. One means the response filled the available context. One means the output budget stopped the response. One means Claude declined to respond. They need different repairs.