The Context Window Is an Attention Span, Not Storage

The single mental-model shift that most improves how you engineer context: stop thinking of the window as a container to fill, and start thinking of it as a limited attention span where placement matters as much as inclusion.

Most people picture a context window as a box. You have so many tokens of space; you fill it with what you need; bigger box means you can fit more. This mental model feels natural and is precisely wrong in ways that lead to poor context engineering. The better model, the one that production practitioners internalise, is that the context window is an attention span. That shift changes everything about how you build.

Why Storage Is the Wrong Metaphor

A storage box treats all its contents equally; a byte at the bottom is as retrievable as a byte at the top. Attention does not work that way. The model attention is unevenly distributed across the window, and simply being present in the context does not guarantee the model will weight a piece of information appropriately. Treating the window as storage leads you to dump information in and assume it will be used. It often will not.

The Lost-in-the-Middle Effect

The most important consequence of the attention model is the well-documented "lost in the middle" effect: information at the very beginning and the very end of the context receives disproportionate attention, while information buried in the middle receives the least. A critical instruction placed in the middle of a long context can be effectively ignored, not because the model cannot see it, but because the model attention is drawn elsewhere.

Placement as a Design Decision

Once you accept the attention model, placement becomes a deliberate design decision. Critical instructions belong near the start or the end. Stable, always-relevant guidance can anchor the beginning. The current plan or the most important immediate context can go at the end, in the high-recency zone. The middle is where you put things that need to be present but are less critical. You are not just deciding what to include; you are choreographing where attention falls.

The Recency Exploit

Some production systems exploit the attention model directly. The Manus todo.md pattern rewrites the agent plan at the end of context on every step, ensuring the strategic picture always sits in the highest-attention recency zone rather than getting lost in the middle as context accumulates. This is attention-aware engineering: deliberately placing the most important information where the model will weight it most heavily.

Building With the Model in Mind

The attention-span model also explains why bigger windows do not solve the context problem and why context rot occurs. A larger window has even more middle for information to get lost in. The practical takeaway is to think like you are managing attention, not filling storage: keep the window focused, place critical information where attention is highest, and never assume that merely including something means the model will use it.