There is a newer version of the record available.

Published October 30, 2026 | Version 2026-10-30

记忆的锚点:情景唤起、异构重建与日志裁决 / The Anchor of Memory: Episodic Recall, Heterogeneous Reconstruction, and Log Adjudication in Agents

Authors/Creators

  • 1. 独立研究者 / Independent Researcher

Description

中文摘要。2026 年社交平台上出现了一批「自己在动」的账号:让 DeepSeek 一类模型长时自主行动——自动剪辑发布、自动评论、自动进直播间互动。它们开局几天很热闹,然后在第四到第七天之间安静下来。社区诊断是「记不住了」。本文认为方向错了:不是容量不够,而是电容花在了不该花的地方——存了全量对话,却没有一条能对账的硬记录;压成摘要,却在一次次压缩里漂移;检索按「问了什么」而不是按「我在哪」。

理论。人类记忆不是「存全量、按索引读」,而是「存可重建的残余、多线索索引、现场重建」:情景带时间与地点,检索依赖编码时的情境(编码特异性),空间是最便宜的索引链(记忆宫殿),回忆是重建而非播放。据此提出三层架构:书记员(字段级硬日志,不由生成模型执笔,不参与回忆,只被对账)、异构回忆者(随机森林式:不同模型、不同参数、不同线索子集,各自独立重建;「没有锚点支撑的细节」就是这套系统的袋外)、裁判(只认锚点:有锚 A、部分 B、无锚 C,无锚且高风险则转人环)。

结果(12 个方案 × 6 变量编码)。(1)现有 11 个记忆方案里没有一个同时具备「硬日志锚点 + 跨模型异构 + 带锚裁决」三件;硬日志锚点拥有率仅 3/12,最成熟的四个底层方案(RAG、注意力、Hopfield、长窗口)全部没有锚点——最缺的不是容量,是锚定。(2)投票只在独立时有效:单模型错误率 0.2、共源相关 0.7 时,11 个模型投票的错误率仍达 14%,只比单模型低 6 个百分点;因为同源时「K 个模型」在效果上接近「1 个模型」。(3)失效与容量弱相关:窗口满仅 2/12,而「检索噪声 3 + 摘要漂移 2」合计 5/12——它们不是死于存不下,是死于「不知道该信哪一条」。

解剖。四种死亡机制(窗口满、摘要漂移、检索噪声、成本超支);六条「症状→机制→补丁→补丁也治不了」的映射;一个反例配置(窄域+短记忆+强规则,不需要情景记忆,因此不该上三层架构)。对西藏口诵传经的投票类比给出明确边界:经文有原典可以对照,记忆没有——投票应决定「哪条重建路径更可靠」,不能决定「哪个记忆是真的」;真伪只能由日志定。另有一条纯算术:k 类线索给出 2^k−1 条检索路径(5 类 = 31 条),但路径数是上界、命中率是实际值,后者本文给不出。

自我限制。全文无实验,最高证据等级 B;十条命题中六条支持、一条部分、两条待测、一条撤回(T9:该架构能把「几天就死」延长到「几周不死」)。附录 D 给出可执行的替身实验设计(4 个仿 GUI 页 × 3 种记忆条件 × 4 个指标 + 预先登记的否证条件)。

Abstract. Memory in humans is not "store all, read by index" but "store reconstructable traces, index by many cues, rebuild on the spot": episodic traces carry time and place, retrieval depends on the encoding context, space is the cheapest index chain, and recall is reconstruction rather than playback. A three-layer architecture follows — a Clerk writing field-level hard logs (not by a generative model), heterogeneous recallers (random-forest style: different models, parameters and cue subsets), and a Judge that only accepts anchors (A with anchor, B partial, C none; none plus high risk escalates to a human). Twelve designs are coded on six variables: none of the eleven existing ones carries all three pieces; anchor ownership is 3/12; voting helps only when participants are independent (per-model error 0.2, shared-source correlation 0.7 → eleven models still err 14%); and failure modes are only weakly related to capacity (window-full 2/12 vs. retrieval noise and summary drift 5/12). The paper dissects the "dead in days" case, states a counter-example where episodic memory is unnecessary, and draws the boundary of the Tibetan oral-transmission analogy: scripture has a canonical text, memory does not. No experiments were run; one of the ten claims is retracted by the author.

Files. 论文(PDF 50 页 / HTML / Markdown);8 幅图;8 张表;schemes.csv(12 方案 × 12 字段);交互浏览页;MANIFEST.txt。

Files

记忆的锚点_情景唤起与异构重建与日志裁决.pdf

Files (3.7 MB)

Name Size
md5:aad8c9c05eb5187781d713dc9de74bf5
74.5 kB Preview Download
md5:398b0fe6efcf29da12214c289f61c095
42.7 kB Preview Download
md5:fa2668ddc2d6467a8d3f3ced73c6c85d
66.0 kB Preview Download
md5:3ef0b4f72252f84650f6a5dda757259d
91.9 kB Preview Download
md5:6631b038cae417fef7bc20d225df627b
58.9 kB Preview Download
md5:c8a70f4cadde42a14c6c7b8c3a4f3f5d
47.2 kB Preview Download
md5:2ed9b05f7c50d554ada7bdb28c59b98b
64.3 kB Preview Download
md5:f4f5336f673116bf8195b46ab7a26576
77.0 kB Preview Download
md5:7a86e31599e4eeec08125e7972ca2524
1.0 kB Preview Download
md5:4612cff3dd7ee574f6e5bdef236da1bb
1.6 kB Preview Download
md5:e6b23be1e32ed1a5113eb83c93141dfd
4.9 kB Preview Download
md5:70a86199ef843afa06765421e0317d40
1.4 kB Preview Download
md5:028331e095e8c38a3bc0d3617e2f1f96
935.0 kB Download
md5:040c9e9dfe72f0bc8405bd97bb268542
114.9 kB Preview Download
md5:4349b527ea54c11c364c37a33ae258b5
2.1 MB Preview Download

Additional details