Short-term Memory

01 · 桌面只有六格，十二张便笺怎么放01 · Six desk slots, twelve notes—what stays?

任务从四步变成十六步，便笺板容量不变。保留全部、丢掉最旧，还是把旧内容压成摘要，会怎样影响 token 与下一步成功率？A task grows from four steps to sixteen while the note board stays fixed. How do keeping everything, evicting old notes, or summarizing them affect tokens and next-step success?

02拉长任务，切换淘汰或摘要策略并观察便笺板、token 与成功率Lengthen the Task, Switch Eviction or Summary, and Watch Notes, Tokens, and Success

记忆账本 · 容量策略memory ledger · capacity policy

状态 AMetric A

—

状态 BMetric B

—

状态 CMetric C

—

所有概率、成本、成功率与输出均为固定种子的教学模拟，不代表真实模型输出。把主变量拖回即可复核结论。All probabilities, costs, success rates, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.

03把刚才的变化拆成 4 个零件Break the Change into Four Parts

Short-term Memory

WORKING

当前工作集Working set

只保留下一步真正需要的事实与中间结果。Keep only facts and intermediate results needed for the next step.

CAPACITY

有限容量Finite capacity

便笺槽与 token 都有上限，任务长度不会让容量自动变大。Note slots and tokens are bounded; task length does not expand capacity.

EVICT

淘汰策略Eviction policy

丢弃最旧或低价值条目，可能损失早期约束。Discard old or low-value entries, risking early constraints.

SUMMARY

摘要策略Summary policy

压缩旧条目保留主旨，但会损失细节。Compress old entries to preserve gist while losing detail.

04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction

MECHANISM

working_memory ≤ capacity; overflow → evict ∨ summarize

任务长度超过六格后，淘汰保持较低 token，摘要保留更多约束；全保留则超窗并降低当前步骤成功率。Past six slots, eviction keeps token use lower while summary preserves more constraints. Keeping all notes overflows the window and lowers next-step success.

任务变长Lengthen便笺持续产生notes accumulate

容量触顶Overflow选择淘汰或摘要evict or summarize

权衡结果Trade offtoken 与成功率联动tokens and success move together

05把任务拉到十六步并保留全部便笺Grow to Sixteen Steps and Keep Every Note

主动制造失败Create a failure

边界实验Boundary experiment

短期记忆不是无限抽屉。全保留会让占用超过 100%，当前步骤丢失关键位置并出现明显失败。Short-term memory is not an infinite drawer. Keeping everything pushes occupancy above 100%, displaces key information, and causes visible failure.

尚未执行。先在主交互中观察正常机制。Not run yet. Observe the normal mechanism in the main interaction first.

✗ 短期记忆就是整个上下文窗口。Short-term memory is the whole context window.→ ✓ 它是受策略管理的当前工作集。It is a policy-managed working set.

✗ 只要删除最旧内容就不会错。Deleting the oldest note is always safe.→ ✓ 早期约束可能仍关键，淘汰要考虑价值。Early constraints may remain critical; eviction needs value signals.

✗ 摘要能无损保存所有细节。A summary preserves every detail.→ ✓ 摘要换容量，会牺牲部分细节与来源。Summarization trades capacity for some detail and provenance.

06 · 一句话带走06 · One line to keep

短期记忆是容量有限的任务便笺板；任务变长后必须淘汰或摘要，并同时观察 token 占用与当前步骤成功率。Short-term memory is a capacity-limited task board. As tasks grow, it must evict or summarize while tracking both token use and next-step success.