TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · The Brief · AGE 18—20 · FACTUAL · EVOLVING
Structuring the brief for attention
A dropped setback rule is often a structure problem, not a memory problem.
The idea
A model holds a limited window of text and does not use every part of it equally well. Research on 2023 models found that material near the start and end was used more reliably than the middle. Newer models have narrowed that gap; how much remains depends on the model and the task. So a brief has a shape, not only a content. Put load-bearing constraints at the top, or in the latest turn, and state them again when the conversation grows long. A constraint stated once, forty exchanges ago, is still in the window and may be ignored.
Why it matters
"The model forgot my setback rule" is often a structure problem, not a memory problem. Once you know that placement matters on your model, the complaint becomes a decision about where to put the rule.
See it in the studio
You set "every plan at 1:100, north up" at the start of a long session. Forty turns later a plan arrives at 1:50, north to the left. Nothing broke. The rule sank into a long middle. Say it again, close to where you need it, and ask whether the long thread is now costing you more than it is keeping for you.
Watch for this
Treating a long thread as a free archive. Every earlier turn dilutes the attention available for the current one. Sometimes the right move is a fresh thread with a tightened brief that carries the constraints forward.
Try it
Take a constraint the model dropped deep in a long chat. Re-anchor it — state it again at the top of your next message — and run the same request. Then open a fresh thread with only the brief and the constraint and run it a third time. Compare all three.
Prove it
Show a brief in which the load-bearing constraints sit where your model uses them best, and explain the placement in terms of the window — and of what you tested, not what you assumed.
How it works
Models read the whole conversation every time they answer, through attention, and attention is a fixed amount of weight spread across everything in the window. The "lost in the middle" pattern was measured on 2023 models. Newer models have softened the curve a great deal — some now retrieve from very long contexts almost perfectly on the tests their makers publish — and products add memory or summaries on top, which changes how forgetting feels. That is why the card is EVOLVING: placement is a model- and task-dependent heuristic, not a law of language models. The habit costs nothing, and the warning against treating a long thread as a free archive holds whatever the curve does.
What this idea builds on
What this idea opens up
- Nothing yet names this as a foundation.
Sources
- prompt-theory-primer
- Liu et al., Lost in the Middle, 2023
- Tokens, Context, and Why AI Forgets
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.