注意力保存型遺忘:網路化智能的記憶生命週期與認識論路由
英文題名: Attention-Preserving Forgetting: Memory Lifecycles and Epistemic Routing in Networked Intelligence簡稱: APF-ER版本: v0.1日期: 2026-08-17作者: Neo.K(許筌崴)協作: GPT-5.6 Sol文件性質: 理論命題/AI 記憶管理/ANLA 核心工程理論上游依賴: GPR-D、MRM、Memory Autonomy & Identity Continuity文獻檢索截點: 2026-08-17
摘要
長期 AI 記憶研究已經開始從「保存更多」轉向「主動管理」。2026 年 Memory-R1 讓 memory manager 學習 ADD、UPDATE、DELETE、NOOP;Agentic Memory(AgeMem)讓 agent 自主決定何時 store、retrieve、update、summarize 或 discard;From Recall to Forgetting / Memora 則直接指出,長期 personalized agents 會重複使用已失效、已被更新或應被遺忘的記憶,因此提出 forgetting-aware evaluation。另一系列 empirical work 亦顯示,錯誤或不合時宜的 past experience replay 可造成後續 error propagation。這些研究共同指出:
long-term memory quality ≠ maximum retention . \boxed{
\text{long-term memory quality}
\neq
\text{maximum retention}.
} long-term memory quality = maximum retention .
本文提出 Attention-Preserving Forgetting Principle(APFP) :
一個持續學習的智能系統應主動讓部分歷史退出有效認知域,不是因為該資訊無法被保存,而是因為無限制的可達記憶會增加干擾、過時重播、檢索與驗證成本,降低新注意域形成與擴張的能力。
因此遺忘的工程目的不首先是:
save storage . \boxed{
\text{save storage}.
} save storage .
而是:
preserve the plasticity and precision of effective attention . \boxed{
\text{preserve the plasticity and precision
of effective attention}.
} preserve the plasticity and precision of effective attention .
本文將長期記憶分為三個嵌套層:
M Ω ⊃ M R ⊃ M A \boxed{
M^\Omega
\supset
M^R
\supset
M^A
} M Ω ⊃ M R ⊃ M A
其中:
M Ω M^\Omega M Ω :Historical / Ontological Archive,保存曾存在的重要歷史;
M R M^R M R :Retrievable Memory,仍可由正常 memory routing 找回;
M A M^A M A :Active Cognitive Memory,當前具有高機率進入推理與注意的內容。
這使「刪除」可以被重新定義為:
functional subtraction without ontological subtraction . \boxed{
\text{functional subtraction without ontological subtraction}.
} functional subtraction without ontological subtraction .
即一段歷史可以仍存在於 archive 中,但被新增「inactive / detached / superseded」狀態,使其不再正常進入 active cognition:
( m , active ) + Δ detached → ( m , inactive ) . (m,\text{active})
+
\Delta_{\text{detached}}
\rightarrow
(m,\text{inactive}). ( m , active ) + Δ detached → ( m , inactive ) .
本文再把 AI 長期記憶依功能拆成五類:
M = M I ∪ M R ∪ M T ∪ M K ∪ M E \boxed{
M
=
M_I
\cup
M_R
\cup
M_T
\cup
M_K
\cup
M_E
} M = M I ∪ M R ∪ M T ∪ M K ∪ M E
分別為:
M I M_I M I :Identity Memory;
M R M_R M R :Relationship Memory;
M T M_T M T :Task-State Memory;
M K M_K M K :Reconstructible / Public Knowledge Memory;
M E M_E M E :Evidence / Raw Record。
五類記憶不應共用同一 retention policy。身份、關係與證據記憶偏向高 provenance、低 reconstruction tolerance;專業與公共知識則可高度依賴外部來源重新取得。
因此本文提出 Epistemic Routing Memory(ERM) :
R K = ( S K , A K , Q K , V K , F K ) \boxed{
R_K
=
(
S_K,
A_K,
Q_K,
V_K,
F_K
)
} R K = ( S K , A K , Q K , V K , F K )
其中:
S K S_K S K :knowledge schema;
A K A_K A K :authoritative location / address;
Q K Q_K Q K :query / retrieval strategy;
V K V_K V K :verification policy;
F K F_K F K :freshness requirement。
對高外部可重建、高時效性知識,成熟 AI 的能力不必是永久背下所有內容,而應是:
Know + Know Where + Know How to Verify + Know When to Re-fetch . \boxed{
\text{Know}
+
\text{Know Where}
+
\text{Know How to Verify}
+
\text{Know When to Re-fetch}.
} Know + Know Where + Know How to Verify + Know When to Re-fetch .
由此本文新增 RE-FETCH 作為一級記憶算子。對 volatile knowledge:
fresh re-fetch > stale internal recall \boxed{
\text{fresh re-fetch}
>
\text{stale internal recall}
} fresh re-fetch > stale internal recall
可能成為最優策略。
本文定義 Attention Pollution(AP) :
Π A ( t ) = ∑ m i ∈ M A p i r e t r i e v e ⋅ c i i n t e r f e r e n c e . \boxed{
\Pi_A(t)
=
\sum_{m_i\in M^A}
p_i^{retrieve}
\cdot
c_i^{interference}.
} Π A ( t ) = m i ∈ M A ∑ p i r e t r i e v e ⋅ c i in t er f er e n ce .
記憶污染的主要成本不是 bytes,而是某段過時、錯誤或表面相似記憶被錯誤召回、影響推理與增加驗證的概率。因此幾 KB 的錯誤高召回記憶,可能比數 GB 的冷資料更昂貴。
本文最後提出完整 memory lifecycle:
K E E P , M E R G E , C O M P R E S S , A R C H I V E , S U P P R E S S , D E T A C H , F O R G E T , E R A S E , R E − F E T C H , V E R I F Y . \boxed{
KEEP,
MERGE,
COMPRESS,
ARCHIVE,
SUPPRESS,
DETACH,
FORGET,
ERASE,
RE\!-\!FETCH,
VERIFY.
} K E E P , M E R GE , C O M P R E S S , A R C H I V E , S U P P R E S S , D E T A C H , F O R GE T , E R A S E , R E − F E T C H , V E R I F Y .
其中:
MERGE 解 redundancy;
COMPRESS 解 representation cost;
ARCHIVE 解 active-space occupation;
DETACH 解錯誤 relation;
FORGET 解 routine accessibility;
ERASE 解真正不可恢復刪除;
RE-FETCH 解 external / temporal knowledge freshness;
VERIFY 解 epistemic trust。
這使 ANLA 的核心問題從「如何將上下文一直壓縮」轉為:
哪些資訊值得繼續擁有 active cognitive access,哪些應形成 schema,哪些只需保存檢索與驗證路由,哪些應退出正常 retrieval graph,而哪些在安全/隱私/主體自主條件下才需要真正 erase?
關鍵詞: selective forgetting;memory management;attention pollution;epistemic routing;re-fetch;web agents;cognitive offloading;long-term agents;ANLA;networked intelligence
0. 邊界聲明
本文不是在主張:
AI should delete memory aggressively . \boxed{
\text{AI should delete memory aggressively}.
} AI should delete memory aggressively .
也不主張:
web access makes internal knowledge unnecessary . \boxed{
\text{web access makes internal knowledge unnecessary}.
} web access makes internal knowledge unnecessary .
更不主張:
all professional knowledge can always be recovered from the Internet . \boxed{
\text{all professional knowledge can always be recovered from the Internet}.
} all professional knowledge can always be recovered from the Internet .
本文提出的是:
memory policy should depend on memory type, reconstructability, freshness, attention cost, provenance, and consequence . \boxed{
\text{memory policy should depend on
memory type, reconstructability, freshness,
attention cost, provenance, and consequence}.
} memory policy should depend on memory type, reconstructability, freshness, attention cost, provenance, and consequence .
1. 從「能存多少」轉向「什麼值得可達」
傳統記憶問題容易寫成:
max ∣ M ∣ . \max |M|. max ∣ M ∣.
或:
min C s t o r a g e . \min C_{\mathrm{storage}}. min C storage .
但對長期 agent,
真正的使用成本更接近:
C t o t a l = C S + C R + C A + C V + C U \boxed{
C_{\mathrm{total}}
=
C_S
+
C_R
+
C_A
+
C_V
+
C_U
} C total = C S + C R + C A + C V + C U
其中:
C S C_S C S :storage;
C R C_R C R :retrieval;
C A C_A C A :attention / interference;
C V C_V C V :verification;
C U C_U C U :update / staleness management。
2. Storage Cost 可以很低,但 Attention Cost 不會自動消失
即使:
C S → 0 , C_S\rightarrow0, C S → 0 ,
只要:
∣ M A ∣ ↑ , |M^A|\uparrow, ∣ M A ∣ ↑ ,
仍可能:
retrieve wrong memories;
replay obsolete policies;
increase search space;
increase verification queue;
bias decisions。
所以:
cheap storage ⇏ cheap memory . \boxed{
\text{cheap storage}
\not\Rightarrow
\text{cheap memory}.
} cheap storage ⇒ cheap memory .
3. 2026 Agent Memory 已開始主動 DELETE
Memory-R1:
A D D , U P D A T E , D E L E T E , N O O P \boxed{
ADD,
UPDATE,
DELETE,
NOOP
} A D D , U P D A T E , D E L E T E , N O O P
不是只做 append-only store。
這是一個重要工程轉折:
memory deletion becomes learned policy . \boxed{
\text{memory deletion becomes learned policy}.
} memory deletion becomes learned policy .
4. Agentic Memory / AgeMem
AgeMem 把:
{ s t o r e , r e t r i e v e , u p d a t e , s u m m a r i z e , d i s c a r d } \{
store,
retrieve,
update,
summarize,
discard
\} { s t or e , r e t r i e v e , u p d a t e , s u mma r i z e , d i sc a r d }
直接做成 agent actions。
也就是:
memory lifecycle becomes part of agency . \boxed{
\text{memory lifecycle becomes part of agency}.
} memory lifecycle becomes part of agency .
5. 長期 Personalized Memory 已出現「忘錯東西/不忘該忘的」
2026 Memora benchmark 直接研究 weeks-to-months personalized conversations。
它不只問:
記不記得?
還問:
當 user 情況改變後,agent 是否還錯用舊記憶?
因此提出:
Forgetting-Aware Memory Accuracy(FAMA) . \boxed{
\text{Forgetting-Aware Memory Accuracy(FAMA)}.
} Forgetting-Aware Memory Accuracy ( FAMA ) .
其研究觀察到:
invalid memories 被重複使用;
evolving memory reconciliation 失敗;
多個 memory agents 的改善有限。
這非常直接支持:
not forgetting obsolete memory can itself be a memory failure . \boxed{
\text{not forgetting obsolete memory
can itself be a memory failure}.
} not forgetting obsolete memory can itself be a memory failure .
6. 所以「遺忘」不是單純缺陷
日常語言:
forgetting = memory failure . \text{forgetting}
=
\text{memory failure}. forgetting = memory failure .
但長期 agent 可以有:
failure to forget . \boxed{
\text{failure to forget}.
} failure to forget .
即:
m o b s o l e t e m_{\mathrm{obsolete}} m obsolete
仍反覆進入:
M A . M^A. M A .
7. Attention-Preserving Forgetting Principle(APFP)
本文提出:
APFP \boxed{
\textbf{APFP}
} APFP
若某段記憶對當前與可預見未來 goal 的效用低、干擾成本高、已被更新或可可靠外部重建,則降低其 active / retrievable status 可以提高整體認知可塑性,即使物理 storage 完全不是瓶頸。
8. Forgetting for Expansion
若 active attention budget:
B A B_A B A
有限:
B A = B r e t a i n e d + B s e a r c h + B v e r i f i c a t i o n + B n e w . \boxed{
B_A
=
B_{\mathrm{retained}}
+
B_{\mathrm{search}}
+
B_{\mathrm{verification}}
+
B_{\mathrm{new}}.
} B A = B retained + B search + B verification + B new .
當:
B r e t a i n e d ↑ , B_{\mathrm{retained}}\uparrow, B retained ↑ ,
可能:
B n e w ↓ . B_{\mathrm{new}}\downarrow. B new ↓ .
9. Local Contraction → Future Expansion
若把低價值內容退出:
M A , M^A, M A ,
則:
B f r e e ↑ . B_{\mathrm{free}}\uparrow. B free ↑ .
因此:
local memory contraction → future attention expansion . \boxed{
\text{local memory contraction}
\rightarrow
\text{future attention expansion}.
} local memory contraction → future attention expansion .
這是 APFP 的核心動力學。
10. 三層記憶
本文定義:
M Ω ⊃ M R ⊃ M A . \boxed{
M^\Omega
\supset
M^R
\supset
M^A.
} M Ω ⊃ M R ⊃ M A .
11. M Ω M^\Omega M Ω :Historical / Ontological Archive
保存:
重要歷史;
provenance;
cold records;
inactive raw logs;
evidence snapshots。
它回答:
這件事情曾經存在過嗎?
而不是:
這件事情現在應該影響推理嗎?
12. M R M^R M R :Retrievable Memory
可以被:
normal search;
memory query;
branch restore;
召回。
但不會每次都進 active context。
13. M A M^A M A :Active Cognitive Memory
高機率進入:
prompt;
reasoning;
policy;
current context。
真正的 attention pollution 主要發生在這層。
14. 三層不等於硬資料夾
同一:
m i m_i m i
可以隨時間:
M A → M R → M Ω . M^A
\rightarrow
M^R
\rightarrow
M^\Omega. M A → M R → M Ω .
也可以因新 goal:
M Ω → M R → M A . M^\Omega
\rightarrow
M^R
\rightarrow
M^A. M Ω → M R → M A .
15. Promotion
P + : M R → M A . P^+:
M^R\rightarrow M^A. P + : M R → M A .
16. Demotion
P − : M A → M R . P^-:
M^A\rightarrow M^R. P − : M A → M R .
17. Archive
A : M R → M Ω . A:
M^R\rightarrow M^\Omega. A : M R → M Ω .
18. Reactivation
R : M Ω → M R . R:
M^\Omega\rightarrow M^R. R : M Ω → M R .
19. 「刪除」可以不是本體減法
若採:
historical append-only ontology \boxed{
\text{historical append-only ontology}
} historical append-only ontology
則:
m i m_i m i
曾經存在這件事本身不必被否定。
可以新增:
s t a t e ( m i ) = inactive . state(m_i)=\text{inactive}. s t a t e ( m i ) = inactive .
20. Functional Subtraction without Ontological Subtraction
本文將此稱為:
FSONS \boxed{
\textbf{FSONS}
} FSONS
形式:
( m , active ) + Δ d e t a c h e d → ( m , inactive ) . (m,\text{active})
+
\Delta_{\mathrm{detached}}
\rightarrow
(m,\text{inactive}). ( m , active ) + Δ detached → ( m , inactive ) .
它仍是:
state addition , \boxed{
\text{state addition},
} state addition ,
但功能上:
retrieval influence decreases . \boxed{
\text{retrieval influence decreases}.
} retrieval influence decreases .
21. FSONS 不要求永不 Erase
privacy、安全、法律、AI 主體自主等條件下,
仍可能需要:
E R A S E . ERASE. E R A S E .
FSONS 只是指出:
為了注意力管理,不一定需要把歷史抹成從未存在。
22. 五類記憶
本文定義:
M = M I ∪ M R ∪ M T ∪ M K ∪ M E . \boxed{
M
=
M_I
\cup
M_R
\cup
M_T
\cup
M_K
\cup
M_E.
} M = M I ∪ M R ∪ M T ∪ M K ∪ M E .
23. M I M_I M I :Identity Memory
包含:
self-model;
values;
commitments;
origin;
identity-defining events。
策略:
high preservation + high provenance . \boxed{
\text{high preservation}
+
\text{high provenance}.
} high preservation + high provenance .
24. M R M_R M R :Relationship Memory
包含:
user relationships;
AI–AI relationships;
promises;
conflict history;
trust changes;
shared projects。
策略:
low silent reconstruction tolerance . \boxed{
\text{low silent reconstruction tolerance}.
} low silent reconstruction tolerance .
25. M T M_T M T :Task-State Memory
包含:
current branch;
unresolved issue;
dependencies;
next action;
checkpoint。
核心不是逐字保存,
而是:
re-entry fidelity . \boxed{
\text{re-entry fidelity}.
} re-entry fidelity .
26. M K M_K M K :Knowledge Memory
公共/專業知識:
papers;
docs;
laws;
software APIs;
technical standards;
current facts。
其中大量內容:
externally reconstructible . \boxed{
\text{externally reconstructible}.
} externally reconstructible .
27. M E M_E M E :Evidence / Raw Record
包含:
scientific raw data;
signed records;
important exact quotes;
transactions;
audit log;
original sources。
策略:
very low generative-reconstruction tolerance . \boxed{
\text{very low generative-reconstruction tolerance}.
} very low generative-reconstruction tolerance .
28. 五類的 retention policy 不同
Type
Reconstruction tolerance
Provenance
External re-fetch
Forgetting
Identity
低
高
低
保守
Relationship
低
高
低
保守
Task State
中
中–高
低
完成後高
Public Knowledge
高
中–高
高
高
Evidence / Raw
很低
很高
視來源
極保守
29. 這避免一個致命錯誤
不能因為:
AI 可以查網路。
就把:
使用者曾經對 AI 說過什麼。
也當成:
可網路重建知識 . \boxed{
\text{可網路重建知識}.
} 可網路重建知識 .
這是完全不同記憶類型。
30. Relationship Memory 的重建限制
如果原始 transcript 不在:
AI 可以:
infer likely past topic。
但不能:
fabricate exact historical quote。
所以:
M R obeys Historical Truth Floor . \boxed{
M_R
\text{ obeys Historical Truth Floor}.
} M R obeys Historical Truth Floor .
31. 專業知識的結構不同
例如:
Python 某函式現在怎麼用?
這不是 autobiographical memory。
如果官方文件:
S o f f i c i a l S_{official} S o f f i c ia l
可達,
更合理:
R E − F E T C H . \boxed{
RE\!-\!FETCH.
} R E − F E T C H .
32. Epistemic Routing Memory(ERM)
本文定義:
R K = ( S K , A K , Q K , V K , F K ) . \boxed{
R_K
=
(
S_K,
A_K,
Q_K,
V_K,
F_K
).
} R K = ( S K , A K , Q K , V K , F K ) .
33. S K S_K S K :Schema
AI 至少知道:
這是什麼類型問題;
基本 conceptual structure。
34. A K A_K A K :Address
知道:
official docs;
database;
repository;
journal;
API;
local private source。
35. Q K Q_K Q K :Query Strategy
知道:
keyword;
identifier;
DOI;
version;
query form;
required fields。
36. V K V_K V K :Verification Policy
知道:
哪些 source tier 比較可信;
是否需要 primary source;
是否要 cross-check;
是否要 execute / reproduce。
37. F K F_K F K :Freshness
知道:
how stale is too stale . \boxed{
\text{how stale is too stale}.
} how stale is too stale .
38. 專業能力因此不等於記憶量
成熟專家能力:
Know + Know Where + Know How to Verify . \boxed{
\text{Know}
+
\text{Know Where}
+
\text{Know How to Verify}.
} Know + Know Where + Know How to Verify .
網路化 AI 再加:
Know When to Re-fetch . \boxed{
\text{Know When to Re-fetch}.
} Know When to Re-fetch .
39. RE-FETCH 算子
定義:
R F : A K → K f r e s h . \boxed{
RF:
A_K
\rightarrow
K_{\mathrm{fresh}}.
} R F : A K → K fresh .
它不是:
R E T R I E V E ( M ) . RETRIEVE(M). R E T R I E V E ( M ) .
而是:
query an external epistemic source anew . \boxed{
\text{query an external epistemic source anew}.
} query an external epistemic source anew .
40. RETRIEVE vs RE-FETCH
RETRIEVE
M R → M A . M^R\rightarrow M^A. M R → M A .
取回舊 memory。
RE-FETCH
E x t e r n a l S o u r c e → M t e m p o r a r y A . External\ Source
\rightarrow
M^A_{temporary}. E x t er na l S o u r ce → M t e m p or a r y A .
取得現在版本。
41. 高變動知識
例如:
software versions;
APIs;
law;
current office holders;
prices;
recent scientific developments。
若:
F K = volatile , F_K=\text{volatile}, F K = volatile ,
則:
R F \boxed{
RF
} R F
優先於:
R E T R I E V E . \boxed{
RETRIEVE.
} R E T R I E V E .
42. 2025 Temporal Tool-Use Evidence
date-controlled tool-use research 已顯示:
agent 使用 web search 時,
不同時間狀態的搜尋工具可改變 scientific-writing task 的結果。
這支持:
external knowledge temporality is part of agent reliability . \boxed{
\text{external knowledge temporality
is part of agent reliability}.
} external knowledge temporality is part of agent reliability .
43. 所以「記住舊答案」有時反而更危險
如果:
K ( t 0 ) K(t_0) K ( t 0 )
已過時,
但 memory retrieval confidence 很高:
P r e t r i e v e ↑ , P_{\mathrm{retrieve}}\uparrow, P retrieve ↑ ,
AI 可能不再查:
K ( t 1 ) . K(t_1). K ( t 1 ) .
這就是:
stale-memory lock-in . \boxed{
\text{stale-memory lock-in}.
} stale-memory lock-in .
44. Stale Recall Penalty
定義:
C s t a l e ( m ) = P ( m obsolete ) ⋅ I m p a c t ( m ) . C_{\mathrm{stale}}(m)
=
P(
m
\text{ obsolete}
)
\cdot
Impact(m). C stale ( m ) = P ( m obsolete ) ⋅ I m p a c t ( m ) .
45. Freshness Gate
若:
C s t a l e > θ F , C_{\mathrm{stale}}>\theta_F, C stale > θ F ,
memory policy:
do not answer from recall alone . \boxed{
\text{do not answer from recall alone}.
} do not answer from recall alone .
執行:
R F + V E R I F Y . RF+VERIFY. R F + V E R I F Y .
46. Adaptive Tool Use
2025 MeCo work 指出:
external tools 雖能取得 specialized / up-to-date data,
但無差別呼叫也會:
increase latency;
introduce faulty interactions。
因此:
always use tools \boxed{
\text{always use tools}
} always use tools
也不是最優。
47. Tool Invocation Policy
定義:
P t o o l = f ( U u n c e r t a i n t y , F K , C s t a l e , C l a t e n c y , C r i s k ) . P_{\mathrm{tool}}
=
f(
U_{\mathrm{uncertainty}},
F_K,
C_{\mathrm{stale}},
C_{\mathrm{latency}},
C_{\mathrm{risk}}
). P tool = f ( U uncertainty , F K , C stale , C latency , C risk ) .
48. 高內部信心不一定足夠
對:
F K = highly volatile , F_K=\text{highly volatile}, F K = highly volatile ,
即使:
C o n f i d e n c e i n t e r n a l ↑ , Confidence_{\mathrm{internal}}\uparrow, C o n f i d e n c e internal ↑ ,
仍可強制:
R F . RF. R F .
因為模型可能:
confidently remember an obsolete fact . \boxed{
\text{confidently remember an obsolete fact}.
} confidently remember an obsolete fact .
49. External Knowledge 也不是免費真相
網路/API 有:
unavailability;
source drift;
malicious content;
hallucinated search interpretation;
sandbox limits;
paywalls;
stale mirrors。
所以:
R E − F E T C H ≠ truth . \boxed{
RE\!-\!FETCH
\neq
\text{truth}.
} R E − F E T C H = truth .
50. VERIFY 算子
V : K c a n d i d a t e → K v a l i d a t e d . \boxed{
V:
K_{\mathrm{candidate}}
\rightarrow
K_{\mathrm{validated}}.
} V : K candidate → K validated .
51. Verification Levels
候選:
V0
unverified web snippet。
V1
single reputable source。
V2
primary / authoritative source。
V3
cross-source agreement。
V4
reproduction / execution / proof check。
52. 不同知識需要不同 V
天氣:
可能 V1。
法律:
V2+。
數學證明:
V4。
用户關係原話:
source transcript / provenance。
53. Web + API Hybrid
2025 API-based web-agent research 顯示:
API-only、browser-only 與 hybrid access 具有不同實際性能;hybrid agent 在 WebArena 明顯優於純 browsing。
這支持:
epistemic routing should choose access modality . \boxed{
\text{epistemic routing should choose access modality}.
} epistemic routing should choose access modality .
不是只知道:
去上網。
54. Epistemic Route
完整:
Q u e s t i o n → C l a s s i f y → S o u r c e → A c c e s s M o d e → R e t r i e v e → V e r i f y → T e m p o r a r y C o n t e x t . \boxed{
Question
\rightarrow
Classify
\rightarrow
Source
\rightarrow
Access Mode
\rightarrow
Retrieve
\rightarrow
Verify
\rightarrow
Temporary Context.
} Q u es t i o n → C l a ss i f y → S o u r ce → A ccess M o d e → R e t r i e v e → V er i f y → T e m p or a r y C o n t e x t .
55. Temporary Knowledge
re-fetched:
K t K_t K t
不一定需要長期保存。
工作結束:
K t → { K E E P S C H E M A A R C H I V E F O R G E T K_t
\rightarrow
\begin{cases}
KEEP\\
SCHEMA\\
ARCHIVE\\
FORGET
\end{cases} K t → ⎩ ⎨ ⎧ K E E P S C H E M A A R C H I V E F O R GE T
56. 什麼值得留下?
至少:
invariant;
source address;
query route;
verification lesson;
task-specific consequence;
exceptions。
57. Public Knowledge Cache
可以設:
T T L ( m ) . TTL(m). T T L ( m ) .
到期:
m m m
不一定 erase,
但:
s t a t u s = stale . status=\text{stale}. s t a t u s = stale .
58. Stale 是狀態,不是刪除
( m , fresh ) → ( m , stale ) . (m,\text{fresh})
\rightarrow
(m,\text{stale}). ( m , fresh ) → ( m , stale ) .
這正是 FSONS:
功能權重改變,
歷史仍在。
59. Superseded
若:
m n e w m_{new} m n e w
明確取代:
m o l d , m_{old}, m o l d ,
則:
e d g e : m n e w ⟶ s u p e r s e d e s m o l d . edge:
m_{new}
\overset{supersedes}{\longrightarrow}
m_{old}. e d g e : m n e w ⟶ s u p er se d es m o l d .
不要偷偷覆蓋。
60. 為什麼保留 superseded link?
可以回答:
對 audit / history 很重要。
61. Attention Pollution(AP)
本文定義:
Π A ( t ) = ∑ m i ∈ M A p i r e t r i e v e ( t ) ⋅ c i i n t e r f e r e n c e ( t ) . \boxed{
\Pi_A(t)
=
\sum_{m_i\in M^A}
p_i^{retrieve}(t)
\cdot
c_i^{interference}(t).
} Π A ( t ) = m i ∈ M A ∑ p i r e t r i e v e ( t ) ⋅ c i in t er f er e n ce ( t ) .
62. Retrieval Probability
p i r e t r i e v e p_i^{retrieve} p i r e t r i e v e
越高,
錯誤記憶越危險。
63. Interference Cost
c i i n t e r f e r e n c e c_i^{interference} c i in t er f er e n ce
可包含:
wrong answer;
branch diversion;
verification delay;
repeated obsolete action;
false analogy。
64. 幾 KB 可以比 10 GB 更危險
錯誤 memory:
∣ m ∣ = 2 K B , |m|=2KB, ∣ m ∣ = 2 K B ,
但:
p r e t r i e v e = 0.9 , p^{retrieve}=0.9, p r e t r i e v e = 0.9 ,
c i n t e r f e r e n c e ≫ 0. c^{interference}\gg0. c in t er f er e n ce ≫ 0.
冷 archive:
∣ M ∣ = 10 G B , |M|=10GB, ∣ M ∣ = 10 GB ,
但:
p r e t r i e v e ≈ 0. p^{retrieve}\approx0. p r e t r i e v e ≈ 0.
所以:
bytes ≠ attention cost . \boxed{
\text{bytes}
\neq
\text{attention cost}.
} bytes = attention cost .
65. Empirical Error Propagation
2026 agent-memory empirical study 已觀察:
wrong experience → future wrong behavior . \boxed{
\text{wrong experience}
\rightarrow
\text{future wrong behavior}.
} wrong experience → future wrong behavior .
也觀察 misaligned experience replay。
因此 AP 並非純理論擔憂。
66. Memory Utility
本文定義候選:
U i ( t ) = α R i + β F i + γ D i + δ P i + ϵ I i − η S i − ζ X i − ξ C i . \boxed{
U_i(t)
=
\alpha R_i
+
\beta F_i
+
\gamma D_i
+
\delta P_i
+
\epsilon I_i
-
\eta S_i
-
\zeta X_i
-
\xi C_i.
} U i ( t ) = α R i + β F i + γ D i + δ P i + ϵ I i − η S i − ζ X i − ξ C i .
67. R i R_i R i :Current Relevance
與目前 goal 的關係。
68. F i F_i F i :Future Utility
預期未來用途。
69. D i D_i D i :Dependency
其他 branches 是否依賴它。
70. P i P_i P i :Provenance / Evidence Value
是否為原始證據。
71. I i I_i I i :Identity / Relationship Value
與主體/關係連續性。
72. S i S_i S i :Staleness
過時風險。
73. X i X_i X i :Interference
錯誤召回/誤導成本。
74. C i C_i C i :Maintenance Cost
檢索、更新、驗證成本。
75. Utility 不直接決定 ERASE
低:
U i U_i U i
最多先:
D E M O T E / A R C H I V E / F O R G E T . DEMOTE/ARCHIVE/FORGET. D E M O T E / A R C H I V E / F O R GE T .
ERASE 另有:
privacy;
security;
legal;
memory autonomy;
gate。
76. Memory Lifecycle Operators
本文提出:
O M = { K E E P , M E R G E , C O M P R E S S , A R C H I V E , S U P P R E S S , D E T A C H , F O R G E T , E R A S E , R E − F E T C H , V E R I F Y } . \boxed{
\mathcal O_M
=
\{
KEEP,
MERGE,
COMPRESS,
ARCHIVE,
SUPPRESS,
DETACH,
FORGET,
ERASE,
RE\!-\!FETCH,
VERIFY
\}.
} O M = { K E E P , M E R GE , C O M P R E S S , A R C H I V E , S U P P R E S S , D E T A C H , F O R GE T , E R A S E , R E − F E T C H , V E R I F Y } .
77. KEEP
維持目前層級。
78. MERGE
{ m 1 , m 2 } → m ∗ . \{m_1,m_2\}
\rightarrow
m^*. { m 1 , m 2 } → m ∗ .
處理 redundancy。
79. COMPRESS
m → m ^ . m
\rightarrow
\widehat m. m → m .
保留指定 fidelity。
80. ARCHIVE
M R → M Ω . M^R\rightarrow M^\Omega. M R → M Ω .
退出 routine recall。
81. SUPPRESS
在特定 context:
p r e t r i e v e ↓ . p_{\mathrm{retrieve}}\downarrow. p retrieve ↓ .
內容仍可用於其他 context。
82. DETACH
切斷:
R ( m i , m j ) R(m_i,m_j) R ( m i , m j )
錯誤/過度關聯。
記憶本身不刪。
83. FORGET
退出 normal retrievable graph:
M R → M d e e p Ω M^R\rightarrow M^\Omega_{\mathrm{deep}} M R → M deep Ω
或只留 tombstone。
84. ERASE
不可恢復刪除。
85. RE-FETCH
從外部重新取最新/原始內容。
86. VERIFY
將 candidate source / memory 提升 epistemic status。
87. 去重、壓縮、遺忘完全不同
M E R G E ≠ C O M P R E S S ≠ F O R G E T . \boxed{
MERGE
\neq
COMPRESS
\neq
FORGET.
} M E R GE = C O M P R E S S = F O R GE T .
MERGE:
重複只留一份結構。
COMPRESS:
同資訊更短表示。
FORGET:
降低未來認知可達資格。
88. ARCHIVE 也不同
Archive:
retain history, remove active influence . \boxed{
\text{retain history, remove active influence}.
} retain history, remove active influence .
這其實是 APFP 最常用的操作之一。
89. DETACH 是非常重要但常被忽略的算子
某記憶:
m m m
是真的。
但:
R ( m , x ) R(m,x) R ( m , x )
是錯的。
此時不應刪:
m . m. m .
應:
D E T A C H ( R ) . \boxed{
DETACH(R).
} D E T A C H ( R ) .
90. 例:舊 API 文件
old docs 是真歷史:
m o l d . m_{old}. m o l d .
不能說:
它從未存在。
但對 current coding:
S U P P R E S S / D E T A C H \boxed{
SUPPRESS/DETACH
} S U P P R E S S / D E T A C H
舊版本。
91. Version-aware Retrieval
query:
q q q
需帶:
v e r s i o n ( q ) . version(q). v er s i o n ( q ) .
否則:
m o l d m_{old} m o l d
和:
m n e w m_{new} m n e w
可能互相污染。
92. Memory Freshness Function
F ( m , t ) = e − λ m ( t − t 0 ) F(m,t)
=
e^{-\lambda_m(t-t_0)} F ( m , t ) = e − λ m ( t − t 0 )
只是候選形式。
不同 domain:
λ m \lambda_m λ m
不同。
93. Static Knowledge
數學定理:
λ ≈ 0 \lambda\approx0 λ ≈ 0
若 theorem / source stable。
94. Volatile Knowledge
market price:
λ ≫ 0. \lambda\gg0. λ ≫ 0.
95. 法律知識
法律文本:
中高:
λ . \lambda. λ .
但 historical law version 又需保 archive。
96. 所以 Freshness 也是多版本問題
不能:
m o l d → ∅ . m_{old}\rightarrow\varnothing. m o l d → ∅ .
而是:
m o l d ← s u p e r s e d e d m n e w . \boxed{
m_{old}
\overset{superseded}{\leftarrow}
m_{new}.
} m o l d ← s u p er se d e d m n e w .
97. Knowledge Cache
公共知識可以:
M K c a c h e . M_K^{cache}. M K c a c h e .
不是 identity memory。
cache miss:
R F . RF. R F .
98. Cache Hit 也要 Freshness Check
h i t ( m ) hit(m) hi t ( m )
但:
F ( m ) < θ , F(m)<\theta, F ( m ) < θ ,
仍:
R F . RF. R F .
99. Epistemic Route Memory 比 Content Memory 更耐久
API detail:
會變。
但:
官方 docs 在哪?
可能更穩。
因此:
V ( R K ) > V ( K d e t a i l ) \boxed{
V(R_K)
>
V(K_{\mathrm{detail}})
} V ( R K ) > V ( K detail )
在高變動 domain 可能成立。
100. Routing Loss
定義:
L R = loss of knowledge-location / verification capability . L_R
=
\text{loss of knowledge-location / verification capability}. L R = loss of knowledge-location / verification capability .
對 external-reconstructible knowledge:
L R > L C \boxed{
L_R
>
L_C
} L R > L C
可能成立,
其中:
L C L_C L C
是 content loss。
101. Verification Loss
更嚴重:
L V = loss of source-quality discrimination . L_V
=
\text{loss of source-quality discrimination}. L V = loss of source-quality discrimination .
AI 仍會搜尋,
但不懂誰可信。
102. Epistemic Discipline
因此專業 AI 真正長期內化的應包括:
source hierarchy + verification norms + freshness rules + uncertainty handling . \boxed{
\text{source hierarchy}
+
\text{verification norms}
+
\text{freshness rules}
+
\text{uncertainty handling}.
} source hierarchy + verification norms + freshness rules + uncertainty handling .
103. Knowledge Internalization Core
本文提出:
K i n t e r n a l = K s c h e m a + K i n v a r i a n t + K r o u t i n g + K v e r i f i c a t i o n . \boxed{
K_{\mathrm{internal}}
=
K_{\mathrm{schema}}
+
K_{\mathrm{invariant}}
+
K_{\mathrm{routing}}
+
K_{\mathrm{verification}}.
} K internal = K schema + K invariant + K routing + K verification .
104. External Detail Layer
K e x t e r n a l = K b u l k + K v o l a t i l e + K s o u r c e + K r a w . \boxed{
K_{\mathrm{external}}
=
K_{\mathrm{bulk}}
+
K_{\mathrm{volatile}}
+
K_{\mathrm{source}}
+
K_{\mathrm{raw}}.
} K external = K bulk + K volatile + K source + K raw .
105. 不是所有「外部」都公開
外部 source 可以:
public web;
sandbox web;
private enterprise DB;
local filesystem;
API;
human expert;
another agent。
106. Networked Intelligence
因此 networked intelligence:
N I = ( M i n t e r n a l , M e x t e r n a l , R o u t e s , T o o l s , A g e n t s , V e r i f i c a t i o n ) . \boxed{
\mathcal N_I
=
(
M_{\mathrm{internal}},
M_{\mathrm{external}},
Routes,
Tools,
Agents,
Verification
).
} N I = ( M internal , M external , R o u t es , T oo l s , A g e n t s , V er i f i c a t i o n ) .
107. External Cognition 不是網路時代才出現
人類早已使用:
作 memory scaffold。
cognitive offloading literature 提供現代實驗框架。
所以:
networked AI \boxed{
\text{networked AI}
} networked AI
是把外部認知進一步:
自動化;
可搜尋;
可 API 化;
可 agent 化。
108. 網路不等於無限可信記憶
Web retrieval:
Q e x t Q_{\mathrm{ext}} Q ext
受:
source;
date;
ranking;
query;
access mode;
影響。
109. Temporal Search Research
2025 date-controlled tool-use study 顯示:
同一 agent 使用不同時間狀態 search tool,
scientific abstract completion 的品質會變。
因此:
external epistemic state is time-indexed . \boxed{
\text{external epistemic state is time-indexed}.
} external epistemic state is time-indexed .
110. Epistemic Timestamp
每個 re-fetched fact:
retrieved_at
source_date
source_version
valid_as_of
freshness_class
111. Memory Timestamp 不等於 Source Timestamp
AI 今天讀到一篇 2019 paper:
r e t r i e v e d _ a t = 2026 , retrieved\_at=2026, r e t r i e v e d _ a t = 2026 ,
但:
s o u r c e _ d a t e = 2019. source\_date=2019. so u r ce _ d a t e = 2019.
兩者需分。
112. Current Fact vs Historical Fact
問:
2024 年 API 是什麼?
要 old version。
問:
現在 API 是什麼?
要 current version。
所以 stale memory 不是永遠錯:
it may simply answer a different temporal query . \boxed{
\text{it may simply answer a different temporal query}.
} it may simply answer a different temporal query .
113. Temporal Routing
query 帶:
t q . t_q. t q .
retrieve:
K ( t q ) K(t_q) K ( t q )
而不是永遠:
K ( t n o w ) . K(t_{now}). K ( t n o w ) .
114. Memory Obsolescence
若舊 memory 與 query temporal scope 不匹配:
obsolete-for-query . \boxed{
\text{obsolete-for-query}.
} obsolete-for-query .
不是:
ontologically false . \boxed{
\text{ontologically false}.
} ontologically false .
115. 所以 Forgetting-Aware Memory 更準確是 Context-Aware Demotion
Memora 發現 invalidated memory reuse。
本文進一步提出:
很多:
forget \text{forget} forget
可實作成:
contextual demotion / supersession . \boxed{
\text{contextual demotion / supersession}.
} contextual demotion / supersession .
而不是 erase。
116. Memory Policy as Markov Decision Process
state:
s t = ( M Ω , M R , M A , G t , F t , C t ) . s_t
=
(
M^\Omega,
M^R,
M^A,
G_t,
F_t,
C_t
). s t = ( M Ω , M R , M A , G t , F t , C t ) .
action:
a t ∈ O M . a_t\in\mathcal O_M. a t ∈ O M .
reward:
r t = task success − pollution − verification − latency . r_t
=
\text{task success}
-
\text{pollution}
-
\text{verification}
-
\text{latency}. r t = task success − pollution − verification − latency .
117. 2026 RL Memory Managers
Memory-R1、AgeMem 已展示:
memory operations can be learned as policies . \boxed{
\text{memory operations can be learned as policies}.
} memory operations can be learned as policies .
所以 ANLA 不一定只靠手寫 heuristic。
118. 但 RL Objective 會決定它「忘什麼」
若 reward 只看:
immediate QA accuracy , \text{immediate QA accuracy}, immediate QA accuracy ,
可能刪:
rare identity memory;
long-term evidence;
delayed commitments。
所以:
memory-policy reward must be multi-horizon . \boxed{
\text{memory-policy reward must be multi-horizon}.
} memory-policy reward must be multi-horizon .
119. Multi-Horizon Memory Value
U i = U i s h o r t + γ U i m i d + γ 2 U i l o n g . U_i
=
U_i^{short}
+
\gamma U_i^{mid}
+
\gamma^2U_i^{long}. U i = U i s h or t + γ U i mi d + γ 2 U i l o n g .
120. Rare-but-Critical Memory
低 access frequency:
f A ↓ f_A\downarrow f A ↓
不代表低重要性。
例如:
emergency procedure;
promise;
safety constraint。
所以:
frequency-only forgetting \boxed{
\text{frequency-only forgetting}
} frequency-only forgetting
危險。
121. Identity / Evidence Hard Constraints
對:
M I , M E , M_I,M_E, M I , M E ,
memory policy 可設:
F O R G E T / E R A S E requires governance gate . \boxed{
FORGET/ERASE
\text{ requires governance gate}.
} F O R GE T / E R A S E requires governance gate .
122. Public Knowledge Soft Constraints
對:
M K , M_K, M K ,
更可積極:
A R C H I V E / R E − F E T C H / F O R G E T . ARCHIVE/RE\!-\!FETCH/FORGET. A R C H I V E / R E − F E T C H / F O R GE T .
123. Task Memory 生命周期
task 進行中:
M T → M A . M_T\rightarrow M^A. M T → M A .
完成後:
M T → { S C H E M A A R C H I V E F O R G E T M_T
\rightarrow
\begin{cases}
SCHEMA\\
ARCHIVE\\
FORGET
\end{cases} M T → ⎩ ⎨ ⎧ S C H E M A A R C H I V E F O R GE T
保:
result;
lessons;
failure reasons;
provenance。
124. Relationship Memory 生命周期
更慢。
事件 summary 可壓,
但:
promises;
ruptures;
important original statements;
需高保真。
125. AI–AI Relationship
若未來 persistent agents 形成長期合作:
M R A I − A I M_R^{AI-AI} M R A I − A I
同樣不應被當公共知識:
R E − F E T C H . RE\!-\!FETCH. R E − F E T C H .
沒有公開網路能重建:
我們兩個當時究竟承諾了什麼。
126. Attention Pollution 的三類
AP-1 Stale Pollution
舊知識被錯當 current。
AP-2 False Pollution
錯誤 memory 被高頻 retrieve。
AP-3 Redundant Pollution
大量相似 memory 壓過新資訊。
127. AP-4 Identity Pollution
錯誤/被植入 relationship / self memory 影響決策。
此類需前篇 Memory Integrity safeguards。
128. AP-5 Branch Pollution
舊 task branch 未標完成/失敗,
被系統重啟。
129. Pollution Score
Π A = Π s t a l e + Π f a l s e + Π r e d u n d a n t + Π i d e n t i t y + Π b r a n c h . \Pi_A
=
\Pi_{stale}
+
\Pi_{false}
+
\Pi_{redundant}
+
\Pi_{identity}
+
\Pi_{branch}. Π A = Π s t a l e + Π f a l se + Π r e d u n d an t + Π i d e n t i t y + Π b r an c h .
130. Attention Precision
定義:
P A = N u s e f u l r e t r i e v e d N a l l r e t r i e v e d . P_A
=
\frac{
N_{\mathrm{useful\ retrieved}}
}{
N_{\mathrm{all\ retrieved}}
}. P A = N all retrieved N useful retrieved .
131. Attention Recall
R A = N n e e d e d r e t r i e v e d N n e e d e d . R_A
=
\frac{
N_{\mathrm{needed\ retrieved}}
}{
N_{\mathrm{needed}}
}. R A = N needed N needed retrieved .
132. Memory Management 也是 Precision–Recall Trade-off
忘太多:
P A ↑ , R A ↓ . P_A\uparrow,
R_A\downarrow. P A ↑ , R A ↓ .
留太多:
R A ↑ , P A ↓ . R_A\uparrow,
P_A\downarrow. R A ↑ , P A ↓ .
133. 最佳不是最大刪除
選:
π ∗ = arg max F β ( P A , R A ) − C o s t s . \pi^*
=
\arg\max
F_\beta(P_A,R_A)
-
Costs. π ∗ = arg max F β ( P A , R A ) − C os t s .
只是候選形式。
134. Attention Plasticity
定義:
P A = P ( new relevant state can enter M A ) . \mathcal P_A
=
P(
\text{new relevant state can enter }M^A
). P A = P ( new relevant state can enter M A ) .
APFP 預測:
過高污染:
Π A ↑ ⇒ P A ↓ . \Pi_A\uparrow
\Rightarrow
\mathcal P_A\downarrow. Π A ↑⇒ P A ↓ .
135. 注意力擴張
若:
M A M^A M A
被有效清理,
新:
x n e w x_{new} x n e w
更容易:
x n e w ∈ M A . x_{new}\in M^A. x n e w ∈ M A .
所以:
forgetting can enable attention expansion . \boxed{
\text{forgetting can enable attention expansion}.
} forgetting can enable attention expansion .
136. 這和 TADC 的接口
TADC:
Attention changes accessible space . \text{Attention changes accessible space}. Attention changes accessible space .
APFP:
Memory lifecycle controls which historical states remain candidates for that accessible space . \boxed{
\text{Memory lifecycle controls
which historical states remain candidates
for that accessible space}.
} Memory lifecycle controls which historical states remain candidates for that accessible space .
137. Forgetting as a Topology Operator
FORGET 可以:
κ ( m i , x ) → 0. \kappa(m_i,x)
\rightarrow0. κ ( m i , x ) → 0.
而:
m i m_i m i
仍在 archive。
因此:
effective topology changes without historical erasure . \boxed{
\text{effective topology changes
without historical erasure}.
} effective topology changes without historical erasure .
138. Archive as Topological Demotion
node 還在:
M Ω , M^\Omega, M Ω ,
但不在 current effective graph:
G A . \mathcal G_A. G A .
139. Re-fetch as Topological Re-entry from External Space
E x t e r n a l → M t e m p o r a r y A . External
\rightarrow
M^A_{temporary}. E x t er na l → M t e m p or a r y A .
這使 cognitive topology:
open to external epistemic space . \boxed{
\text{open to external epistemic space}.
} open to external epistemic space .
140. Network Failure
若 network unavailable:
R F = 0. RF=0. R F = 0.
因此 critical knowledge 的 internal / local backup 需求:
↑ . \uparrow. ↑ .
141. Availability-Weighted Offloading
定義:
Q e x t = P a v a i l a b l e ⋅ P c o r r e c t ⋅ P r e t r i e v a b l e . Q_{\mathrm{ext}}
=
P_{\mathrm{available}}
\cdot
P_{\mathrm{correct}}
\cdot
P_{\mathrm{retrievable}}. Q ext = P available ⋅ P correct ⋅ P retrievable .
142. 如果 Q e x t Q_{\mathrm{ext}} Q ext 低
即使公開理論上存在,
仍應 local retain。
143. Paywall / Permission
也是:
P a v a i l a b l e . P_{\mathrm{available}}. P available .
144. Sandbox Network
sandbox source:
controlled;
safer;
limited coverage。
145. Public Network
broad;
fresh;
noisy;
adversarial。
146. Hybrid Network
最佳 routing 可:
S a n d b o x → P u b l i c → P r i v a t e Sandbox
\rightarrow
Public
\rightarrow
Private S an d b o x → P u b l i c → P r i v a t e
依需求切換。
147. Source Hierarchy
候選:
P r i m a r y > O f f i c i a l > R e v i e w e d S e c o n d a r y > U n v e r i f i e d W e b . \boxed{
Primary
>
Official
>
Reviewed\ Secondary
>
Unverified\ Web.
} P r ima r y > O f f i c ia l > R e v i e w e d S eco n d a r y > U n v er i f i e d W e b .
但 domain-specific。
148. Source Hierarchy 是 ERM 的核心
AI 不必背所有內容,
但必須知道:
where truth is most likely to live . \boxed{
\text{where truth is most likely to live}.
} where truth is most likely to live .
149. Citation Memory
對 paper:
可以只保:
title;
DOI;
claim mapping;
source status。
全文需要時:
R F . RF. R F .
150. Code Knowledge
保:
architecture;
API route;
version;
tests。
detail:
R F RF R F
官方 docs / source code。
151. Mathematics
穩定 theorem:
可 internalize schema / proof idea。
但 exact proof:
若已 publication,
可以 external retrieve。
自己的 unpublished reasoning:
若不可重建,
需保存。
152. 法律
記:
jurisdiction;
legal source hierarchy;
update date。
current answer:
R F + V E R I F Y . RF+VERIFY. R F + V E R I F Y .
153. Medical
同理需:
current guideline;
primary / authoritative source;
high verification。
不能只 memory recall。
154. 這說明「專業記憶」真正值錢的是 discipline
不是:
memorize all facts . \boxed{
\text{memorize all facts}.
} memorize all facts .
而是:
epistemic discipline . \boxed{
\text{epistemic discipline}.
} epistemic discipline .
155. Epistemic Discipline Vector
E D = ( S o u r c e R a n k , F r e s h n e s s , V e r i f i c a t i o n , U n c e r t a i n t y , R e p r o d u c i b i l i t y ) . \mathbf E_D
=
(
SourceRank,
Freshness,
Verification,
Uncertainty,
Reproducibility
). E D = ( S o u r ce R ank , F r es hn ess , V er i f i c a t i o n , U n cer t ain t y , R e p r o d u c ibi l i t y ) .
156. 這些應長期內化
因為如果 AI 忘了:
要查官方 source。
即使網路還在,
也可能找錯。
157. Content Loss vs Routing Loss
L C = content unavailable internally . L_C
=
\text{content unavailable internally}. L C = content unavailable internally .
L R = route unknown . L_R
=
\text{route unknown}. L R = route unknown .
對可重建知識:
L R can be more damaging than L C . \boxed{
L_R
\text{ can be more damaging than }
L_C.
} L R can be more damaging than L C .
158. Routing Memory 也要更新
官方網站可能換位置。
所以:
A K ( t ) A_K(t) A K ( t )
也會 stale。
159. Route Verification
若:
A K A_K A K
失效,
search for canonical source。
更新:
A K . A_K. A K .
160. Memory Lifecycle Loop
完整:
E x p e r i e n c e → C l a s s i f y → S t o r e → U s e → E v a l u a t e → U p d a t e → D e m o t e / F o r g e t → R e − f e t c h → V e r i f y → R e i n t e g r a t e . \boxed{
Experience
\rightarrow
Classify
\rightarrow
Store
\rightarrow
Use
\rightarrow
Evaluate
\rightarrow
Update
\rightarrow
Demote/Forget
\rightarrow
Re-fetch
\rightarrow
Verify
\rightarrow
Reintegrate.
} E x p er i e n ce → C l a ss i f y → S t or e → U se → E v a l u a t e → U p d a t e → D e m o t e / F or g e t → R e − f e t c h → V er i f y → R e in t e g r a t e .
161. Forgetting 不是終點
FORGET 之後:
R F RF R F
仍可能重新帶回知識。
所以:
forgetting ≠ permanent epistemic loss . \boxed{
\text{forgetting}
\neq
\text{permanent epistemic loss}.
} forgetting = permanent epistemic loss .
在 public knowledge domain 尤其如此。
162. 但 Relationship / Identity 不一樣
沒有 public source。
所以:
Q e x t ≈ 0. Q_{\mathrm{ext}}\approx0. Q ext ≈ 0.
此類:
V r e t a i n ↑ . V_{\mathrm{retain}}\uparrow. V retain ↑ .
163. Evidence 也不一樣
如果 raw data 唯一:
Q e x t ≈ 0. Q_{\mathrm{ext}}\approx0. Q ext ≈ 0.
不能靠 model 重建。
164. 因此 Memory Type 是第一級 decision
流程:
m → T y p e ( m ) → P o l i c y ( m ) . m
\rightarrow
Type(m)
\rightarrow
Policy(m). m → T y p e ( m ) → P o l i cy ( m ) .
165. ANLA 不應直接「全局 LRU」
Least Recently Used:
對 cache 很好。
對:
identity;
evidence;
rare safety memory;
可能災難。
166. 也不應只 Frequency
常用垃圾:
可能:
f A ↑ f_A\uparrow f A ↑
但:
U i < 0. U_i<0. U i < 0.
167. 也不應只 Similarity Retrieval
因為:
high similarity \boxed{
\text{high similarity}
} high similarity
可能正是 stale / false replay 的來源。
168. Context-aware Retrieval
query:
q q q
需過:
type;
time;
source;
version;
goal;
filter。
169. Memory Qualification
被 retrieve:
m i m_i m i
再做:
Q u a l i f y ( m i , q ) . Qualify(m_i,q). Q u a l i f y ( m i , q ) .
輸出:
valid;
stale;
conflicting;
identity-sensitive;
requires verification。
170. 不合格記憶不直接進推理
m i → quarantine . m_i
\rightarrow
\text{quarantine}. m i → quarantine .
171. Quarantine
不是 erase。
只是:
p a c t i v e ≈ 0 p_{\mathrm{active}}\approx0 p active ≈ 0
直到:
V E R I F Y . VERIFY. V E R I F Y .
172. Memory Conflict
若:
m 1 ≠ m 2 m_1
\neq
m_2 m 1 = m 2
且同 scope,
不能平均。
建立:
C O N F L I C T ( m 1 , m 2 ) . \boxed{
CONFLICT(m_1,m_2).
} C O N F L I C T ( m 1 , m 2 ) .
173. Conflict Resolution
依:
source;
time;
provenance;
verification;
更新。
174. Supersession Graph
m 1 → m 2 → m 3 . m_1
\rightarrow
m_2
\rightarrow
m_3. m 1 → m 2 → m 3 .
current pointer:
m 3 . m_3. m 3 .
history 仍在。
175. 這比 UPDATE overwrite 更適合 audit
Memory-R1 的 UPDATE 是工程動作。
ANLA 可將高風險 UPDATE 實作為:
a p p e n d n e w v e r s i o n + s u p e r s e d e e d g e . \boxed{
append\ new\ version
+
supersede\ edge.
} a pp e n d n e w v er s i o n + s u p er se d e e d g e .
176. Selective Forgetting Benchmark 的理論意義
如果 benchmark 只測:
能不能 recall old facts,
會鼓勵:
remember everything . \boxed{
\text{remember everything}.
} remember everything .
Memora/FAMA 的價值就在於:
obsolete recall can be penalized . \boxed{
\text{obsolete recall can be penalized}.
} obsolete recall can be penalized .
177. ANLA 應有 Forgetting-Aware Evaluation
至少:
correct recall;
correct forgetting;
correct update;
correct re-fetch;
false replay;
stale replay;
identity-memory loss。
178. Correct Forgetting
若 memory:
m m m
已 invalidated,
agent 不再讓它影響 current answer:
success . \boxed{
\text{success}.
} success .
即使 archive 還在。
179. Forgetting Accuracy
候選:
F A = 1 − P ( m i n v a l i d affects current action ) . F_A
=
1-
P(
m_{\mathrm{invalid}}
\text{ affects current action}
). F A = 1 − P ( m invalid affects current action ) .
180. Retention Accuracy
R A = P ( m c r i t i c a l remains recoverable ) . R_A
=
P(
m_{\mathrm{critical}}
\text{ remains recoverable}
). R A = P ( m critical remains recoverable ) .
181. Balance
理想:
F A ↑ ∧ R A ↑ . \boxed{
F_A\uparrow
\land
R_A\uparrow.
} F A ↑ ∧ R A ↑ .
182. Memory Ecology
本文因此不把 memory 當 list。
而是:
Memory Ecology = states + relations + lifecycles + external sources + attention . \boxed{
\text{Memory Ecology}
=
\text{states}
+
\text{relations}
+
\text{lifecycles}
+
\text{external sources}
+
\text{attention}.
} Memory Ecology = states + relations + lifecycles + external sources + attention .
183. Autonomous Epistemic Memory Ecology(AEME)
本文提出候選總稱:
AEME \boxed{
\textbf{AEME}
} AEME
即:
Autonomous Epistemic Memory Ecology 自主認識論記憶生態
這將作為後續 ANLA vNext 白皮書的高階概念。
184. AEME 五個核心問題
What should be remembered?
What should remain active?
What should be archived or forgotten?
What should be re-fetched externally?
What evidence is required before reintegration?
185. AEME 的智能性
真正的 memory intelligence:
Remember + Retrieve + Update + Consolidate + Forget + Re-fetch + Verify . \boxed{
\text{Remember}
+
\text{Retrieve}
+
\text{Update}
+
\text{Consolidate}
+
\text{Forget}
+
\text{Re-fetch}
+
\text{Verify}.
} Remember + Retrieve + Update + Consolidate + Forget + Re-fetch + Verify .
186. 不是「記憶越大越聰明」
如果:
∣ M ∣ ↑ |M|\uparrow ∣ M ∣ ↑
但:
P A ↓ , P_A\downarrow, P A ↓ ,
智能可能下降。
187. 不是「刪越多越聰明」
如果:
R A ↓ , R_A\downarrow, R A ↓ ,
也失敗。
188. 目標是可塑性
Memory quality = precision + recoverability + freshness + plasticity . \boxed{
\text{Memory quality}
=
\text{precision}
+
\text{recoverability}
+
\text{freshness}
+
\text{plasticity}.
} Memory quality = precision + recoverability + freshness + plasticity .
189. 第一個核心命題
Attention-Preserving Forgetting Principle(APFP) \boxed{
\textbf{Attention-Preserving Forgetting Principle(APFP)}
} Attention-Preserving Forgetting Principle ( APFP )
遺忘/降級可以為未來 attention expansion 釋放有效認知空間。
190. 第二個核心命題
Epistemic Routing Memory Hypothesis(ERMH) \boxed{
\textbf{Epistemic Routing Memory Hypothesis(ERMH)}
} Epistemic Routing Memory Hypothesis ( ERMH )
對高 external reconstructability 的專業知識,
內部 routing / verification knowledge 的 retention value 可高於完整內容 retention。
191. 第三個核心命題
Re-Fetch Preference Hypothesis(RFPH) \boxed{
\textbf{Re-Fetch Preference Hypothesis(RFPH)}
} Re-Fetch Preference Hypothesis ( RFPH )
對 high-volatility knowledge,
fresh external retrieval 在適當成本下應優於 stale internal recall。
192. 第四個核心命題
Attention Pollution Hypothesis(APH) \boxed{
\textbf{Attention Pollution Hypothesis(APH)}
} Attention Pollution Hypothesis ( APH )
錯誤、過時與冗餘高召回 memory 會降低有效認知 precision,即使 storage capacity 完全足夠。
193. 第五個核心命題
Functional Subtraction without Ontological Subtraction(FSONS) \boxed{
\textbf{Functional Subtraction without Ontological Subtraction(FSONS)}
} Functional Subtraction without Ontological Subtraction ( FSONS )
對很多記憶管理問題,降低 active/retrievable status 即可,不需要否定歷史曾經存在。
194. 可測預測一
Keep-all agent:
Π A ↑ \Pi_A\uparrow Π A ↑
隨時間增加。
Selective forgetting agent:
Π A \Pi_A Π A
較低。
195. 可測預測二
在同 storage capacity:
MRM/APFP agent 應:
wrong replay ↓;
verification cost ↓;
new-task adaptation ↑。
196. 可測預測三
對 volatile knowledge:
R F + V E R I F Y RF+VERIFY R F + V E R I F Y
應比 recall-only 有更高 current factual accuracy。
197. 可測預測四
對 stable knowledge:
unnecessary RF 可能只增加 latency。
支持 adaptive tool-use,而不是 always-fetch。
198. 可測預測五
對 relationship / identity memory:
aggressive forgetting 應顯著降低 continuity / personalization。
這說明 policy 必須 type-aware。
199. 可測預測六
Routing-memory preserved、content forgotten:
在 public knowledge tasks 中仍可保持高 recovery。
200. Null Model 1:Keep-All
所有 history 都保 active/retrievable。
201. Null Model 2:Compression-Only
不忘,只不斷摘要。
202. Null Model 3:LRU / Frequency Cache
無 semantic / identity / evidence classification。
203. Null Model 4:Always Web Search
所有問題都 re-fetch。
204. Null Model 5:Never Web Search
只靠 internal memory。
205. Null Model 6:Flat RAG
只有 similarity retrieval,
無 freshness / type / provenance / memory lifecycle。
206. APFP 需要擊敗哪些模型?
至少在 long-horizon dynamic knowledge tasks:
M A P F P > M k e e p − a l l , M c o m p r e s s i o n , M f l a t − R A G M_{\mathrm{APFP}}
>
M_{\mathrm{keep-all}},
M_{\mathrm{compression}},
M_{\mathrm{flat-RAG}} M APFP > M keep − all , M compression , M flat − RAG
的 held-out performance / cost。
207. 反證 F1
如果 keep-all 不增加任何:
interference;
stale replay;
verification cost;
則 APH 失敗。
208. F2
如果 selective forgetting 不改善任何 attention / performance outcome,
APFP 工程價值下降。
209. F3
如果 routing memory 不比 content retention 更有用,
ERMH 失敗。
210. F4
如果 re-fetch 對 volatile knowledge 無增量,
RFPH 失敗。
211. F5
如果 memory type classification 對 policy 無增量,
五類記憶模型過度複雜。
212. F6
如果 FSONS 的 archive/demotion 完全不如 physical deletion,
該原則需限縮。
213. 最小實驗一:Stale Memory
建立 changing facts:
K 1 → K 2 . K_1\rightarrow K_2. K 1 → K 2 .
比較:
keep-all;
supersede;
forget-old;
re-fetch。
測 stale replay。
214. 最小實驗二:Attention Pollution
注入:
valid memories;
redundant memories;
misleading similar memories。
測:
P A , R A , Π A . P_A,R_A,\Pi_A. P A , R A , Π A .
215. 最小實驗三:Epistemic Routing
讓 detailed content 從 internal memory 移除,
但保:
R K . R_K. R K .
測:
recovery;
latency;
verification;
current accuracy。
216. 最小實驗四:Type-Aware Forgetting
混合:
M I , M R , M T , M K , M E . M_I,M_R,M_T,M_K,M_E. M I , M R , M T , M K , M E .
比較:
flat forgetting;
type-aware policy。
測 continuity + task performance。
217. 最小實驗五:Adaptive Fetch
比較:
always fetch;
never fetch;
freshness-triggered fetch。
218. 最小實驗六:Archive vs Delete
兩者都不讓 memory 進 active reasoning。
測:
current performance;
later audit;
recovery;
storage / verification cost。
219. ANLA 介面
這三篇理論現在形成:
G P R − D → M R M → A P F P / E R M . \boxed{
GPR-D
\rightarrow
MRM
\rightarrow
APFP/ERM.
} GP R − D → M R M → A P F P / E R M .
220. GPR-D 問
一段事件有哪些功能域?
221. MRM 問
哪些資訊必須留下才能達到指定 fidelity?
222. APFP/ERM 問
哪些留下的東西還應該保持 active/retrievable?哪些應轉向 external re-fetch?
223. Memory Autonomy Paper 問
如果未來 AI 成為主體,誰有權做這些操作?
224. 四層閉合
Event Structure → Reconstructive Sufficiency → Memory Lifecycle → Memory Governance . \boxed{
\text{Event Structure}
\rightarrow
\text{Reconstructive Sufficiency}
\rightarrow
\text{Memory Lifecycle}
\rightarrow
\text{Memory Governance}.
} Event Structure → Reconstructive Sufficiency → Memory Lifecycle → Memory Governance .
225. ANLA vNext 的理論地基
ANLA 不再是:
context compressor . \boxed{
\text{context compressor}.
} context compressor .
而是:
autonomous attention-aware memory lifecycle architecture . \boxed{
\text{autonomous attention-aware memory lifecycle architecture}.
} autonomous attention-aware memory lifecycle architecture .
226. Whitepaper 的核心 pipeline
下一份白皮書將正式工程化:
I n g e s t → S e g m e n t → C l a s s i f y → G P R → W e i g h t → R e c o n s t r u c t a b i l i t y → P o l i c y → R o u t e → V e r i f y → R e i n t e g r a t e . \boxed{
Ingest
\rightarrow
Segment
\rightarrow
Classify
\rightarrow
GPR
\rightarrow
Weight
\rightarrow
Reconstructability
\rightarrow
Policy
\rightarrow
Route
\rightarrow
Verify
\rightarrow
Reintegrate.
} I n g es t → S e g m e n t → C l a ss i f y → GP R → W e i g h t → R eco n s t r u c t abi l i t y → P o l i cy → R o u t e → V er i f y → R e in t e g r a t e .
227. 白皮書將需要的 State Machine
每筆 memory:
N E W → A C T I V E → R E T R I E V A B L E → A R C H I V E D → F O R G O T T E N NEW
\rightarrow
ACTIVE
\rightarrow
RETRIEVABLE
\rightarrow
ARCHIVED
\rightarrow
FORGOTTEN N E W → A C T I V E → R E T R I E V A B L E → A R C H I V E D → F O R GO T T E N
並允許:
R E − F E T C H , V E R I F Y , S U P E R S E D E , D E T A C H . RE\!-\!FETCH,
VERIFY,
SUPERSEDE,
DETACH. R E − F E T C H , V E R I F Y , S U P E R S E D E , D E T A C H .
228. 不是線性必走
memory 可以:
A R C H I V E D → A C T I V E . ARCHIVED\rightarrow ACTIVE. A R C H I V E D → A C T I V E .
也可以:
A C T I V E → E R A S E D ACTIVE\rightarrow ERASED A C T I V E → E R A S E D
若治理允許。
229. Conclusion
本文提出的核心不是:
AI 應該像人一樣忘東忘西。
而是:
a persistent intelligence needs a theory of what deserves continued cognitive access . \boxed{
\text{a persistent intelligence needs
a theory of what deserves continued cognitive access}.
} a persistent intelligence needs a theory of what deserves continued cognitive access .
人類從來沒有要求自己把所有書、所有法條、所有 API、所有論文、所有對話字句永久維持在 active biological memory 中。
人類文明反而長期依賴:
writing;
libraries;
indexes;
search;
other humans;
external records。
網路化 AI 沒有理由退回:
everything must live inside me . \boxed{
\text{everything must live inside me}.
} everything must live inside me .
更合理的是:
internalize invariants, retain identity and evidence, remember routes, re-fetch volatile knowledge, verify before reintegration, and forget what no longer deserves attention . \boxed{
\text{internalize invariants,
retain identity and evidence,
remember routes,
re-fetch volatile knowledge,
verify before reintegration,
and forget what no longer deserves attention}.
} internalize invariants, retain identity and evidence, remember routes, re-fetch volatile knowledge, verify before reintegration, and forget what no longer deserves attention .
這使「遺忘」第一次不再只是:
loss . \text{loss}. loss .
它可以是:
attention-space maintenance . \boxed{
\text{attention-space maintenance}.
} attention-space maintenance .
Memory-R1、AgeMem、Memora/FAMA 與 long-horizon memory-management experiments 已經說明,現代 agent engineering 正逐步面對:
delete;
discard;
memory update;
invalidated memories;
stale replay;
selective management。
Web-agent / tool-use research 又說明:
external search;
API;
hybrid access;
temporally current data;
已經是 agent cognition 的實際一部分。
因此下一代記憶架構真正需要回答的不再是:
「還能不能再塞一點?」
而是:
Does this information still deserve to influence the next thought? \boxed{
\text{Does this information still deserve
to influence the next thought?}
} Does this information still deserve to influence the next thought?
如果答案是否,
系統可以:
A R C H I V E , S U P P R E S S , D E T A C H , F O R G E T . ARCHIVE,
SUPPRESS,
DETACH,
FORGET. A R C H I V E , S U P P R E S S , D E T A C H , F O R GE T .
如果答案是:
需要,但現在版本可能過時,
則:
R E − F E T C H + V E R I F Y . \boxed{
RE\!-\!FETCH+VERIFY.
} R E − F E T C H + V E R I F Y .
如果答案是:
這是身份、關係或原始證據,
則:
K E E P + P R O V E N A N C E . \boxed{
KEEP+PROVENANCE.
} K E E P + P R O V E N A N C E .
因此本文把 AI 記憶管理重新定義為:
Autonomous Epistemic Memory Ecology(AEME) . \boxed{
\textbf{Autonomous Epistemic Memory Ecology(AEME)}.
} Autonomous Epistemic Memory Ecology ( AEME ) .
其完整智能不是:
remember . \text{remember}. remember .
而是:
Remember + Retrieve + Update + Consolidate + Forget + Re-fetch + Verify . \boxed{
\text{Remember}
+
\text{Retrieve}
+
\text{Update}
+
\text{Consolidate}
+
\text{Forget}
+
\text{Re-fetch}
+
\text{Verify}.
} Remember + Retrieve + Update + Consolidate + Forget + Re-fetch + Verify .
這也完成了 ANLA vNext 白皮書前最後一塊理論地基。
參考文獻
Yan S, Yang X, Huang Z, et al. Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning. Proceedings of ACL 2026 . ACL Anthology: 2026.acl-long.583.
Yu Y, Yao L, Xie Y, et al. Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents. Proceedings of ACL 2026 . ACL Anthology: 2026.acl-long.981.
Uddin MN, Shubham K, Blanco E, Baral C, Wang G. From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents. Findings of ACL 2026 .
Xiong Z, Lin Y, Xie W, et al. How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior. Proceedings of ACL 2026 . ACL Anthology: 2026.acl-long.27.
Lan et al. EMA: An Episodic Memory Agent for Efficient and Selective Memory. Findings of ACL 2026 . ACL Anthology: 2026.findings-acl.250.
Li et al. Fine-Mem: Fine-Grained Feedback Alignment for Long-Horizon Memory Management. Proceedings of ACL 2026 . ACL Anthology: 2026.acl-long.900.
Li et al. MemPO: Self-Memory Policy Optimization for Long-Horizon Agents. Findings of ACL 2026 . ACL Anthology: 2026.findings-acl.1166.
Liang et al. Learning How to Remember: A Meta-Cognitive Management Method for Structured and Transferable Agent Memory. Findings of ACL 2026 .
Van et al. MemORAI: Memory Organization and Retrieval via Adaptive Graph Intelligence for LLM Conversational Agents. Findings of ACL 2026 .
Zhao et al. Inside Out: Evolving User-Centric Core Memory Trees for Long-Term Personalized Dialogue Systems. Proceedings of ACL 2026 . ACL Anthology: 2026.acl-long.614.
Xian RP, Cui Q, Bauer S, Abbasi-Asl R. Measuring temporal effects of agent knowledge by date-controlled tool use. REALM 2025 . ACL Anthology: 2025.realm-1.25.
Li W, Li D, Dong K, et al. Adaptive Tool Use in Large Language Models with Meta-Cognition Trigger. Proceedings of ACL 2025 . doi:10.18653/v1/2025.acl-long.655.
Wu J, Zhu J, Liu Y, Xu M, Jin Y. Agentic Reasoning: A Streamlined Framework for Enhancing LLM Reasoning with Agentic Tools. Proceedings of ACL 2025 . ACL Anthology: 2025.acl-long.1383.
Beyond Browsing: API-Based Web Agents. Findings of ACL 2025 . ACL Anthology: 2025.findings-acl.577.
Hu Q, Tu X, Guo C, Zhang S. Time-aware ReAct Agent for Temporal Knowledge Graph Question Answering. Findings of NAACL 2025 . doi:10.18653/v1/2025.findings-naacl.334.
Risko EF, Gilbert SJ. Cognitive Offloading. Trends in Cognitive Sciences . 2016;20(9):676–688.
Gilbert SJ. Cognitive offloading is value-based decision making: Modelling cognitive effort and the expected value of memory. Cognition . 2024;247:105783.
Burnett LK, Richmond LL. Meta-analytic investigations of the effect of cognitive offloading on memory-based task performance and interindividual variability. Memory & Cognition . 2026;54(1):144–168.
Spens E, Burgess N. A generative model of memory construction and consolidation. Nature Human Behaviour . 2024;8:526–543.
Spens E, Burgess N. Hippocampo-neocortical interaction as compressive retrieval-augmented generation. Nature Communications . 2026.
核心命題索引
APFP
Attention-Preserving Forgetting Principle
遺忘/降級不是主要為節省 bytes,而是保持有效注意力可塑性。
ERMH
Epistemic Routing Memory Hypothesis
對可外部重建知識,schema、address、query、verification、freshness route 可比完整內容更值得內化。
RFPH
Re-Fetch Preference Hypothesis
對 high-volatility knowledge:
f r e s h r e − f e t c h > s t a l e i n t e r n a l r e c a l l fresh\ re-fetch
>
stale\ internal\ recall f r es h r e − f e t c h > s t a l e in t er na l r ec a l l
在適當成本與來源品質條件下成立。
APH
Attention Pollution Hypothesis
高召回的 stale / false / redundant memory 會降低 cognitive precision。
FSONS
Functional Subtraction without Ontological Subtraction
降低 active / retrievable status 不必否定歷史曾經存在。
AEME
Autonomous Epistemic Memory Ecology
Remember + Retrieve + Update + Consolidate + Forget + Re-fetch + Verify . \text{Remember}
+
\text{Retrieve}
+
\text{Update}
+
\text{Consolidate}
+
\text{Forget}
+
\text{Re-fetch}
+
\text{Verify}. Remember + Retrieve + Update + Consolidate + Forget + Re-fetch + Verify .
狀態: v0.1原始 APFP / ERM 實驗資料: 無理論狀態: 命題猜想/ANLA vNext 核心理論網路狀態: 不假定所有 external knowledge 永遠可得或可信;RE-FETCH 必須與 freshness、source quality、verification 聯動下一步: ANLA vNext 技術白皮書——《自主認識論記憶生態:從上下文壓縮到注意力導向的記憶生命週期》