PART 01
单位是 token Measured in tokens
窗口按 token 计数，不按汉字、单词或文件页数计数。 Capacity is counted in tokens, not characters, words, or document pages.
01 · 想象一张固定宽度的桌子
01 · Picture a desk with fixed width
新便签不断推上桌，桌面满了以后，最早那张还在眼前吗？ As new notes keep arriving, what happens to the oldest one when the desk is full?
02 窗口传送带：逐张塞入 token Window Conveyor: Feed Tokens One at a Time
容量 ≠ 记忆 capacity ≠ memory
窗口内
Visible
0/6
tokens
已掉出
Dropped
0
不再可见
no longer visible
注意力工作量
Attention work
0
n2 pairs
03 把“窗口”拆成 4 个事实 Four Facts Hidden Inside “Window”
讲容量，不讲选材 capacity, not curationPART 02
范围有上限 The range is bounded
超限信息必须被截断、压缩或拒绝，不能假装仍在本次计算里。 Overflow must be truncated, compressed, or rejected; it cannot remain in the same computation by wishful thinking.
PART 03
更大有成本 Larger has a cost
本页用 n2 对比注意力配对量；这是教学尺度，不代表具体模型账单。 This lesson uses n2 attention pairs for intuition; it is not a bill for any specific model.
PART 04
更大不等于更可靠 Larger is not more reliable
无关 token 也会占位置。选择放什么属于 S6 上下文工程，而不是本页职责。 Irrelevant tokens also take space. Choosing what belongs is context engineering in S6, not this page’s job.
04 最小机制：只保留末尾 C 张 Minimal Mechanism: Keep Only the Last C Tokens
sliding window
visible = tokensmax(0,n−C)…n
n 是已送入总数，C
是窗口容量。拖动容量滑块即可验证可见切片怎样改变。
n is the number fed and C is capacity. Move the capacity slider
to verify how the visible slice changes.
新 token 进入
New token enters
总数 n 增加
total n grows
超过容量 C
n exceeds C
最左侧被截去
left edge is trimmed
本次计算只见切片
Only the slice is visible
掉出不等于已记住
dropped does not mean remembered
05 边界实验：用噪声挤掉原始任务 Boundary Test: Push the Task Out with Noise
主动制造失败 Create a failure让“只读”从左侧掉出去 Make “read only” fall off the left edge
固定容量为 4，再塞入整条序列。原始任务会离开窗口；这证明窗口不是永久记忆。 Set capacity to four and feed the full stream. The original task leaves the window, proving that a context window is not permanent memory.
尚未执行。原始任务 token 仍未掉出。
Not run yet. The task token has not fallen out.
✗ 进过窗口的信息就会长期记住。
Anything that entered the window becomes long-term
memory.
→ ✓ 窗口只描述本次计算可见范围。
The window only describes visibility in this
computation.
✗ 窗口越大，答案一定越可靠。
A larger window always makes answers more reliable.
→ ✓ 更大也会容纳更多噪声并增加工作量。
More capacity can hold more noise and require more work.
✗ 窗口页应决定哪些资料值得放。
A window lesson should decide which material belongs.
→ ✓ S3 讲窗口是什么；S6 讲往里放什么。
S3 explains the window; S6 explains what to put inside.
06 · 一句话带走
06 · One line to keep