TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 16—18 · FACTUAL · EVOLVING
The context window is a budget
Every model has a token limit, and early instructions can get buried.
The idea
Every model has a limit to how much information it can use at once. That limit is its context window, measured in tokens. A bigger window does not mean every part gets equal attention. In long sessions, earlier instructions can weaken or get buried. Treat context as working space, not storage: keep what matters visible, and bring things back when you need them.
Why it matters
"The model forgot my project" usually means an instruction fell out of use, not that the tool is broken. Knowing this turns a complaint into a fix.
See it in the studio
You set a rule at the start of a long session: keep every plan at 1:100. Forty exchanges later, a drawing arrives at 1:50. The rule did not break. It sank. Say it again, close to where you need it.
Watch for this
Filling the window because it is big. A larger window is a ceiling to stay under, not a space to fill.
Try it
Put the same instruction near the beginning, the middle, and the end of three otherwise identical long prompts. Compare what survives. You have just discovered what researchers call "lost in the middle" — before reading the paper.
Prove it
Explain, in your own words, why a rule stated once at the start can quietly stop working later — and what you would do about it.
How it works
Models read your whole conversation every time they answer, through a mechanism called attention. Attention is spread across everything in the window, and it does not spread evenly. Research on long contexts has measured that material placed in the middle is recalled less reliably than material at the start or the end. Products also differ: some add memory, retrieval, or summaries on top of the raw window, which changes how forgetting feels.
Lineage
Early chatbots held a few thousand tokens; frontier models now hold hundreds of thousands. The budget grew a hundredfold. The lesson did not change: attention is finite, wherever the ceiling sits.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.