Context Window Management — References

Why the window is bounded at all, what it costs to serve, and the evidence that its usable fraction is smaller than its advertised size.

Why there is a limit

• Attention Is All You Need (Vaswani et al., 2017), arXiv:1706.03762 — the architecture whose attention computation scales quadratically in sequence length, which is the origin of the constraint everything here is managing.
• Efficient Memory Management for Large Language Model Serving with PagedAttention (Kwon et al., 2023), arXiv:2309.06180 — what an occupied window physically costs on the serving side, and why prefix sharing and eviction behave the way they do from a caller's point of view.
• Press, Smith and Lewis, 2021, on attention with linear biases and input-length extrapolation — the clearest treatment of why a model trained at one length does not simply work at a longer one. Cited by name.

Why the usable window is smaller than the advertised one

• Lost in the Middle: How Language Models Use Long Contexts (Liu et al., 2023), arXiv:2307.03172 — measured degradation in the use of material placed away from the beginning and end of a long context, which is the basis for planning against an effective length rather than the datasheet number.
• Hsieh and colleagues, 2024, on measuring the real context size of long-context language models — synthetic tasks at varying depths and lengths, and the demonstration that accepted length and effective length diverge sharply. Cited by name; the arXiv path was not verified.

No verified reference is recorded for session rollover practice, hand-off artefact design, or occupancy telemetry. These are operational patterns; the published work measures what models do with a full window rather than how a product arranges never to have one.