為什麼 AI 代理需要本體論——以及用圖儲存它們¶
文章資訊
作者：Alexander Shereshevsky 日期：2026-06-13
原文標題：Why AI Agents Need Ontologies — and Graphs to Store Them
Medium 原連結：https://medium.com/graph-praxis/why-ai-agents-need-ontologies-and-graphs-to-store-them-b02bc24dbb73
📝 重點摘要¶
TL;DR¶
AI 代理的記憶瓶頸，需靠「本體論定義結構＋圖儲存」雙管齊下解決。
核心問題¶
建構 AI 代理的真正瓶頸不是模型大小或 context 長度，而是記憶。多數團隊用向量資料庫「全部 embed、靠相似度檢索」，做 demo 可以，但無法在長期、可靠的場景下運作。文章主張需要「定義良好的本體論」與「圖式儲存」共同解決此問題。
關鍵發現 / 數據¶
- 引用香港理工大學 Yang et al. (2026) 涵蓋 200+ 篇研究 的 graph-based agent memory 綜述作為論據。
- 區分 knowledge memory（本體論骨架） 與 experience memory（動態實例），兩者缺一不可，後者沒有前者就只是 log。
- 記憶生命週期分四階段：extraction、storage、retrieval、evolution；本體論結構對「演化」階段助益最大。
- 圖記憶分類：知識圖（三元組）、階層圖、時序圖（四元組 + valid_until）、超圖（多實體單一 hyperedge）。
- 時序系統如 Graphiti 維護建立/失效時間戳，靠 temporal invalidation 解決矛盾而非覆寫。
方法亮點¶
- 本體論定義「資料的形狀」而非資料本身——typed entities、named relationships、constraints。
- 檢索從「相似度猜測」變為「結構化導航」：語意錨點 → 結構化遍歷 → 規則過濾 → multi-hop 推理。
- 混合架構：用知識圖存靜態本體知識、用時序/階層圖存動態經驗，透過共享 entity reference 連接。
- 細粒度演化：加邊、調權重、合併重複實體、抽象成高階 schema、修剪過時連結。
對我的研究有用嗎？¶
高度相關。「knowledge vs. experience memory」的二分、以及把 retrieval 視為 policy-based planning 問題的觀點，對 GraphRAG 的 schema 設計與多跳檢索很有參考價值。文中對時序圖（四元組、Graphiti 式失效機制）與超圖處理 n 元關係的描述，可直接啟發知識圖譜的時間建模與複雜關係表達。提到的 associative inference（發掘隱含連結）也呼應 GNN-on-KG 方向。
評語¶
觀點清晰、論證自洽，適合作為立場引導文「快讀」；但屬意見性 blog 而非實證——全文僅轉述單一綜述、無自家 benchmark 或數據，「3-4x 準確率」之類硬證據缺席，且作者立場明顯偏向圖方案，需對照原始 survey 再判斷。
🌐 中英對照¶
I'll output the bilingual Markdown directly instead.
Why AI Agents Need Ontologies — and Graphs to Store Them / 為什麼 AI 智能體需要本體論——以及用來儲存它們的圖¶
Author: Alexander Shereshevsky
Published:
Source: https://medium.com/graph-praxis/why-ai-agents-need-ontologies-and-graphs-to-store-them-b02bc24dbb73
Fetched: 2026-06-13T00:55:54.797006
Why AI Agents Need Ontologies — and Graphs to Store Them / 為什麼 AI 智能體需要本體論——以及用來儲存它們的圖¶
Press enter or click to view image in full size
按下 Enter 或點擊以檢視完整尺寸的圖片
The next leap for autonomous AI isn’t a bigger context window. It’s giving agents a structured understanding of their world — and a memory architecture that can hold it.
自主 AI 的下一次飛躍並不在於更大的上下文視窗 (context window)。而在於賦予智能體 (agent) 對其世界的結構化理解——以及一個能夠承載這種理解的記憶體架構。
A quiet consensus is forming among people building AI agents: memory is the bottleneck. Not model size, not context length, not tool access — memory. The ability to accumulate knowledge, organize it meaningfully, and recall the right pieces at the right time.
在建構 AI 智能體的人群中，一個低調的共識正在形成：記憶才是瓶頸。不是模型大小，不是上下文長度，也不是工具存取權限——而是記憶。也就是累積知識、有意義地組織知識，並在正確的時機回想起正確片段的能力。
Most teams try to solve this with vector databases. Embed everything, retrieve by similarity, hope for the best. It works for demos. It breaks for anything that needs to operate reliably over time.
大多數團隊試圖用向量資料庫 (vector database) 來解決這個問題。把所有東西都做嵌入 (embedding)、按相似度檢索，然後祈求一切順利。這對於展示 (demo) 來說行得通，但對任何需要長時間可靠運作的場景就會失靈。
We’ve been down this road ourselves. And the conclusion we’ve reached — one now backed by a growing body of research, including a recent survey from Hong Kong Polytechnic University covering over 200 works — is that the solution requires two things working together: defined ontologies that give structure to what the agent knows, and graph-based storage that can faithfully represent, query, and evolve that structure over time.
我們自己也走過這條路。而我們得出的結論——如今已有越來越多的研究支持，包括香港理工大學 (Hong Kong Polytechnic University) 近期一篇涵蓋超過 200 篇著作的綜述 (survey)——是：解決方案需要兩件事協同運作：為智能體所知之事賦予結構的已定義本體論 (defined ontologies)，以及能夠隨時間忠實表示、查詢並演化該結構的基於圖的儲存 (graph-based storage)。
These aren’t two separate ideas. They’re two halves of the same insight.
這並不是兩個各自獨立的想法。它們是同一個洞見的兩個半邊。
The Missing Layer: Why Agents Need Ontologies / 缺失的一層：為什麼智能體需要本體論¶
An ontology, in this context, is a formal definition of the types of things that exist in the agent’s world and how they can relate to each other. It’s a schema for understanding — a declaration of “these are the kinds of entities I care about, these are the relationships between them, and these are the constraints that govern them.”
在這個脈絡下，本體論 (ontology) 是對智能體世界中存在的事物類型，以及它們彼此之間如何關聯的一種形式化定義。它是一種理解用的綱要 (schema)——是一份宣告：「這些是我所關心的實體類型，這些是它們之間的關係，而這些是約束它們的限制條件。」
Think of a customer-support agent. Without an ontology, the agent accumulates a flat soup of conversation snippets, tool outputs, and user messages. With an ontology, it knows that a Customer can have Subscriptions, that Subscriptionshave Plans and Billing Cycles, that Issues are linked to Products and have Resolution States, and that a customer’s Sentiment can shift across interactions. The ontology doesn’t contain the data — it defines the shape of the data.
想想一個客戶支援智能體。沒有本體論時，智能體累積的是一鍋扁平的雜燴：對話片段、工具輸出與使用者訊息。有了本體論，它就知道一個客戶 (Customer) 可以擁有訂閱 (Subscriptions)，訂閱擁有方案 (Plans) 與計費週期 (Billing Cycles)，問題 (Issues) 與產品 (Products) 相連並具有解決狀態 (Resolution States)，而且客戶的情緒 (Sentiment) 會在不同互動之間轉變。本體論並不包含資料——它定義的是資料的形狀。
This matters because structure is what makes memory useful. Consider the difference between these two approaches to storing a simple fact — “the user switched from the Pro plan to the Enterprise plan last Tuesday”:
這之所以重要，是因為結構正是讓記憶變得有用的關鍵。想想以下兩種儲存一個簡單事實——「使用者在上週二從 Pro 方案切換到 Enterprise 方案」——的方法之間的差異：
Without an ontology, this becomes a text chunk in a vector store. To retrieve it later, you need a query that happens to be semantically similar. “What plan is the user on?” might work. “What changed recently?” might not surface it at all. “Does this user qualify for the enterprise SLA?” has no chance.
沒有本體論時，這會變成向量庫 (vector store) 中的一段文字區塊 (chunk)。若日後想檢索它，你需要一個剛好在語意上相似的查詢。「使用者用的是哪個方案？」也許行得通。「最近有什麼變動？」也許根本不會浮現出來。「這位使用者是否符合企業級 SLA 的資格？」則毫無機會。
With an ontology, this becomes a structured event: a Plan Change connecting a Customer to a new Subscription with a timestamp. Now any query that touches customers, plans, subscriptions, billing, or recent changes can reach this fact through explicit relationships — not through the accident of embedding similarity.
有了本體論，這會變成一個結構化事件：一筆帶有時間戳記的方案變更 (Plan Change)，將一位客戶連接到一筆新的訂閱。如今任何觸及客戶、方案、訂閱、計費或近期變動的查詢，都能透過明確的關係抵達這個事實——而不必依賴嵌入相似度的偶然性。
The survey from Yang et al. frames this as the distinction between knowledge memory and experience memory. Knowledge memory — the agent’s stable understanding of how the world is organized — is essentially ontological. It provides the scaffolding. Experience memory — the dynamic record of what actually happened — fills that scaffolding with specific instances. Both are necessary. But without the ontological layer, experience memory is just a log.
Yang 等人的綜述將此框定為知識記憶 (knowledge memory) 與經驗記憶 (experience memory) 之間的區別。知識記憶——智能體對世界如何組織的穩定理解——本質上是本體論性的。它提供了鷹架 (scaffolding)。經驗記憶——對實際發生之事的動態紀錄——則以具體的實例填滿那座鷹架。兩者都不可或缺。但若沒有本體論這一層，經驗記憶就只是一份日誌而已。
This is, in our view, the most underappreciated idea in agent architecture today. The AI community has invested enormous energy in retrieval techniques, embedding models, and context management. Far less attention has gone to the prior question: what is the structure of what we’re trying to remember?
在我們看來，這是當今智能體架構中最被低估的想法。AI 社群在檢索技術、嵌入模型與上下文管理上投入了巨大的精力，卻對一個更前置的問題給予遠少得多的關注：我們試圖記住之事的結構究竟是什麼？
Why Graphs Are the Natural Home for Structured Memory / 為什麼圖是結構化記憶的天然歸宿¶
Once you commit to a structured, ontological approach to agent memory, the question of storage becomes almost self-answering. You need a data structure that can represent typed entities, named relationships between them, hierarchical organization, temporal dynamics, and flexible schema evolution. That’s a graph.
一旦你決心採用一種結構化、本體論式的方法來處理智能體記憶，儲存的問題幾乎就不言自明了。你需要一種資料結構，能夠表示帶型別的實體、它們之間具名的關係、層級式組織、時間動態，以及彈性的綱要演化。那就是圖 (graph)。
The survey catalogs several flavors of graph-based memory, and each maps naturally to a different aspect of ontological structure.
這篇綜述編列了幾種基於圖的記憶風格，而每一種都自然對應到本體論結構的某個不同面向。
Knowledge graphs are the most direct expression of ontological memory. They store information as entity-relationship triples — (Customer, subscribedTo, EnterprisePlan) — where the types and relationships are defined by the ontology. This is the agent’s factual backbone: structured, queryable, and interpretable.
知識圖譜 (Knowledge graphs) 是本體論記憶最直接的表達形式。它們以實體—關係三元組 (triple)——(Customer, subscribedTo, EnterprisePlan)——的形式儲存資訊，其中型別與關係由本體論定義。這是智能體的事實骨幹：結構化、可查詢且可解讀。
Hierarchical graphs express the containment and abstraction relationships that ontologies naturally define. A Department contains Teams, which contain Members. A Project has Phases, which have Tasks. These parent-child structures enable the agent to reason at different levels of granularity — zooming out for the big picture, zooming in for specific details.
層級圖 (Hierarchical graphs) 表達了本體論自然定義的包含與抽象關係。一個部門 (Department) 包含團隊 (Teams)，而團隊包含成員 (Members)。一個專案 (Project) 有若干階段 (Phases)，而階段有若干任務 (Tasks)。這些親子結構讓智能體能夠在不同的粒度層級上推理——放大看全局，縮小聚焦於具體細節。
Temporal graphs address a limitation that pure ontologies don’t solve on their own: things change. By extending triples into quadruples — (Customer, subscribedTo, ProPlan, valid_until: Tuesday) — temporal graphs track the evolution of the agent’s world. The survey highlights systems like Graphiti that maintain distinct creation and expiration timestamps, resolving contradictions through temporal invalidation rather than overwriting.
時序圖 (Temporal graphs) 解決了純粹本體論本身無法解決的一個限制：事物會變化。透過將三元組擴展為四元組 (quadruple)——(Customer, subscribedTo, ProPlan, valid_until: Tuesday)——時序圖能追蹤智能體世界的演化。綜述特別提到像 Graphiti 這樣的系統，它們維護各自獨立的建立與到期時間戳記，透過時序失效 (temporal invalidation) 而非覆寫來化解矛盾。
Hypergraphs handle the cases where ontological relationships span more than two entities. A clinical interaction between a Patient, a Medication, and a Condition with a specific Outcome is naturally a single hyperedge, not three separate binary relationships decomposed and stripped of their joint meaning.
超圖 (Hypergraphs) 處理本體論關係橫跨兩個以上實體的情況。一次涉及病患 (Patient)、藥物 (Medication) 與病況 (Condition) 並帶有特定結果 (Outcome) 的臨床互動，天然就是單一條超邊 (hyperedge)，而不是三個分開的二元關係被拆解、並剝除其聯合意義之後的結果。
Most production systems combine several of these into hybrid architectures. The survey describes patterns that separate static ontological knowledge from dynamic experiential data, using knowledge graphs for the former and temporal or hierarchical structures for the latter, connected through shared entity references.
大多數正式上線的系統會將其中幾種結合為混合式架構。綜述描述了一些將靜態本體論知識與動態經驗資料分開的模式：前者使用知識圖譜，後者使用時序或層級結構，並透過共享的實體引用 (entity reference) 加以連結。
The critical advantage of graphs over flat storage isn’t just structural fidelity — it’s retrieval quality. When memory is organized according to an ontology and stored in a graph, retrieval stops being a guessing game based on similarity and becomes structured navigation.
圖相對於扁平儲存的關鍵優勢，不僅在於結構保真度——更在於檢索品質。當記憶依照本體論加以組織並儲存於圖中時，檢索就不再是基於相似度的猜謎遊戲，而成為結構化的導覽。
From Schema to Search: How Structure Transforms Retrieval / 從綱要到搜尋：結構如何改造檢索¶
This is where the two ideas — ontologies and graphs — compound. Retrieval over an ontology-informed graph is categorically different from retrieval over a vector store.
正是在這裡，兩個想法——本體論與圖——產生了複利效應。在一個由本體論指引的圖上進行檢索，與在向量庫上進行檢索，是本質上截然不同的兩回事。
Semantic retrieval remains a starting point. You can embed a query and find a relevant anchor node. But from there, structured traversal takes over. The agent follows typed relationships along paths defined by the ontology — from a customer to their subscription, from a subscription to its billing history, from billing to a dispute, from a dispute to its resolution. Each hop is meaningful because the ontology defines what kinds of connections are valid and relevant.
語意檢索 (Semantic retrieval) 仍是一個起點。你可以將查詢做嵌入，並找到一個相關的錨點節點 (anchor node)。但從那裡開始，便由結構化遍歷 (structured traversal) 接手。智能體沿著本體論所定義的路徑，循著帶型別的關係前進——從一位客戶到他的訂閱，從訂閱到其計費歷史，從計費到一筆爭議，從爭議到其解決方案。每一跳 (hop) 都是有意義的，因為本體論定義了哪些種類的連接是有效且相關的。
Rule-based filtering becomes possible because the ontology defines constraints. You can enforce temporal windows (“only retrieve events from the last 30 days”), type restrictions (“only retrieve entities of type Issue”), or relational constraints (“only retrieve plans connected to this customer”) — all of which are impossible with pure vector search.
基於規則的過濾 (Rule-based filtering) 成為可能，因為本體論定義了約束條件。你可以強制施加時間視窗（「只檢索最近 30 天的事件」）、型別限制（「只檢索型別為 Issue 的實體」），或關係約束（「只檢索與這位客戶相連的方案」）——這些在純向量搜尋中全都是不可能做到的。
Multi-hop reasoning — the ability to chain facts — is the clearest win. The survey describes graph-based operators that expand retrieval outward from anchor nodes, pulling in connected context along relational edges. This is what lets an agent answer questions like “why did our pipeline proposal fail?” by traversing from the proposal to the client’s tech stack to a specific database limitation — a chain of reasoning that vector search simply cannot reproduce.
多跳推理 (Multi-hop reasoning)——將事實串連起來的能力——是最明顯的勝利。綜述描述了一些基於圖的運算子 (operator)，它們從錨點節點向外擴展檢索，沿著關係邊 (relational edge) 拉入相連的上下文。正是這一點讓智能體能夠回答諸如「為什麼我們的管線提案失敗了？」這類問題：從提案遍歷到客戶的技術堆疊，再到某個特定的資料庫限制——這是一條向量搜尋根本無法重現的推理鏈。
The survey also describes policy-based retrieval, where the retrieval process itself is treated as a planning problem. An agentic retriever navigates the graph, deciding which paths to explore based on learned policies. In an ontologically structured graph, these policies can be far more effective because the search space is organized and meaningful rather than flat and arbitrary.
綜述還描述了基於策略的檢索 (policy-based retrieval)，其中檢索過程本身被視為一個規劃問題。一個具備智能體性質的檢索器在圖中導覽，依據習得的策略 (policy) 決定要探索哪些路徑。在一個以本體論結構化的圖中，這些策略能夠有效得多，因為搜尋空間是經過組織且有意義的，而非扁平且任意的。
Memory That Evolves: The Lifecycle Advantage / 會演化的記憶：生命週期的優勢¶
A graph structured by an ontology isn’t just better for storage and retrieval — it’s better for evolution. And memory evolution is arguably the feature that separates genuine long-term memory from sophisticated caching.
一個由本體論構築的圖，不僅在儲存與檢索上更出色——它在演化 (evolution) 上也更出色。而記憶演化，可說正是區分真正的長期記憶與精巧快取 (caching) 的那項特徵。
The survey describes the memory lifecycle as four stages: extraction, storage, retrieval, and evolution. The first three are improved by ontological structure, but it’s evolution where the impact is most profound.
綜述將記憶的生命週期 (lifecycle) 描述為四個階段：抽取 (extraction)、儲存 (storage)、檢索 (retrieval) 與演化 (evolution)。前三者皆因本體論結構而獲得改善，但影響最為深遠之處正是演化。
Internal self-evolving — the agent reorganizing its own memory — works fundamentally differently when memory has ontological structure. Instead of crude operations like “replace this text chunk with a newer one,” graph-based evolution can perform fine-grained updates: adding new edges, modifying relationship weights, merging duplicate entities, abstracting repeated patterns into higher-level schemas, or pruning connections that have been superseded. The ontology provides the rules for what constitutes a valid update. Without it, you’re doing coarse-grained text replacement. With it, you’re performing principled knowledge refinement.
內部自我演化 (Internal self-evolving)——智能體重新組織自身的記憶——在記憶具有本體論結構時，其運作方式根本不同。基於圖的演化不必再進行像「用較新的文字區塊取代這一段」這類粗糙的操作，而能執行細粒度的更新：新增邊、修改關係權重、合併重複實體、將重複出現的模式抽象成更高層級的綱要，或修剪已被取代的連接。本體論提供了什麼才算一次有效更新的規則。沒有它，你做的是粗粒度的文字取代；有了它，你進行的是有原則的知識精煉。
External self-exploration — updating memory based on environmental feedback — also benefits. When an action fails, an ontology-structured memory can pinpoint what in its knowledge was wrong and update the specific relationship or entity, rather than rewriting an entire text chunk and hoping the correction propagates.
外部自我探索 (External self-exploration)——根據環境回饋來更新記憶——同樣受益。當一個動作失敗時，一個由本體論構築的記憶能夠精準指出其知識中哪一處出了錯，並更新那個特定的關係或實體，而不必重寫整段文字並寄望修正能自行擴散。
The survey describes systems that perform associative inference — discovering latent connections between nodes that weren’t explicitly encoded. This is essentially the agent “thinking” by exploring its own memory graph, following paths that the ontology makes meaningful. A flat vector store has no mechanism for this kind of structured introspection.
綜述描述了一些執行聯想推理 (associative inference) 的系統——也就是發現節點之間那些未被明確編碼的潛在連接。這本質上就是智能體透過探索自己的記憶圖、循著本體論使之有意義的路徑來「思考」。一個扁平的向量庫並沒有任何機制能進行這種結構化的內省。
Why This Convergence Matters Now / 為什麼這種匯流在當下至關重要¶
Three forces are making the ontology-plus-graph approach increasingly necessary.
有三股力量正使得「本體論加圖」這套方法日益不可或缺。
Long-horizon agents are becoming real. We’re past the era of single-turn assistants. Agents now manage multi-week projects, maintain ongoing client relationships, and accumulate domain expertise across thousands of interactions. At this scale, unstructured memory isn’t just suboptimal — it’s unworkable. You need the organizational discipline that ontologies provide and the structural fidelity that graphs deliver.
長時程智能體 (Long-horizon agents) 正在成為現實。 我們已經走出單輪 (single-turn) 助理的時代。如今智能體要管理長達數週的專案、維繫持續進行的客戶關係，並在數以千計的互動中累積領域專業知識。在這種規模下，非結構化的記憶不只是次優——而是根本行不通。你需要本體論所提供的組織紀律，以及圖所帶來的結構保真度。
Multi-agent coordination requires shared schemas. When multiple agents collaborate, they need to share memory. Sharing a vector store is possible but chaotic — there’s no agreement on what kinds of information exist or how they relate. Shared ontologies solve this: they provide a common vocabulary that different agents can read, write, and reason over. Graphs make that shared vocabulary queryable.
多智能體協調需要共享的綱要。 當多個智能體協作時，它們需要共享記憶。共享一個向量庫是可行的，但混亂不堪——對於存在哪些種類的資訊、它們之間如何關聯，並沒有任何共識。共享本體論解決了這一點：它們提供了一套共通的詞彙，讓不同的智能體可以讀取、寫入並在其上推理。而圖讓這套共享詞彙變得可查詢。
Trust demands interpretability. In regulated domains — healthcare, finance, legal — users and auditors need to understand why an agent made a decision. Graph-based memory grounded in an explicit ontology provides a traceable chain of reasoning: “I recommended X because of relationship Y between entities A and B, recorded at time T.” Vector similarity scores offer no comparable explanation.
信任要求可解讀性 (interpretability)。 在受監管的領域——醫療、金融、法律——使用者與稽核人員需要理解智能體為何做出某項決策。建立在明確本體論之上的基於圖的記憶，提供了一條可追溯的推理鏈：「我之所以推薦 X，是因為實體 A 與 B 之間在時間 T 所記錄的關係 Y。」向量相似度分數則無法提供任何可與之相比的解釋。
The survey reinforces this by documenting applications across conversational agents, code assistants, financial systems, robotics, scientific discovery, and gaming — all domains where structured, long-term memory is moving from “nice to have” to essential infrastructure.
綜述透過記錄橫跨對話式智能體、程式碼助理、金融系統、機器人學、科學發現與遊戲等領域的應用來強化這一點——在所有這些領域中，結構化的長期記憶正從「有了更好」轉變為不可或缺的基礎設施。
The Hard Problems Ahead / 前方的難題¶
We don’t want to oversell this. Ontology-driven graph memory introduces its own challenges.
我們並不想過度吹捧。由本體論驅動的圖記憶也帶來了它自身的挑戰。
Schema design is hard. Defining the right ontology for a domain requires deep understanding of both the domain and how the agent will use the information. Too rigid, and the schema can’t accommodate unexpected information. Too loose, and you lose the structural benefits. Dynamic schema evolution — letting the ontology itself adapt over time — is an active research frontier.
綱要設計很難。 為某個領域定義出正確的本體論，需要對該領域以及智能體將如何使用這些資訊都有深入的理解。太過僵硬，綱要就容納不下意料之外的資訊；太過鬆散，你就喪失了結構帶來的好處。動態綱要演化——讓本體論本身隨時間自我調適——是一個活躍的研究前沿。
Scalability needs work. Graph operations can be computationally expensive at scale, and the field still needs better solutions for incremental updates, approximate retrieval, and distributed graph storage.
可擴展性 (Scalability) 仍需努力。 圖運算在大規模時可能在計算上代價高昂，而這個領域在增量更新、近似檢索與分散式圖儲存方面仍需要更好的解決方案。
Privacy gets complicated. Relational structures can leak information through inference. Even if individual facts are anonymized, the pattern of connections in a graph can reveal sensitive information. The survey flags this as an open challenge, and we agree — it’s one of the harder problems to solve well.
隱私變得棘手。 關係結構可能透過推理而洩漏資訊。即使個別事實已被匿名化，圖中連接的模式仍可能揭露敏感資訊。綜述將此標示為一個開放性挑戰，我們也同意——這是較難妥善解決的問題之一。
Evaluation is immature. How do you measure whether a memory system’s ontology is “good”? Beyond downstream task accuracy, we need metrics for structural integrity, temporal consistency, and semantic completeness. These don’t exist in standardized form yet.
評估尚不成熟。 你要如何衡量一個記憶系統的本體論是否「良好」？除了下游任務的準確度之外，我們還需要衡量結構完整性、時序一致性與語意完備性的指標。這些目前都還沒有標準化的形式。
Where We Stand / 我們的立場¶
If you’re building AI agents that need to work reliably over time, the evidence is increasingly clear: you need structure, and you need graphs.
如果你正在建構需要長時間可靠運作的 AI 智能體，證據已越來越清楚：你需要結構，而且你需要圖。
The ontology defines the shape of what your agent understands. The graph gives that shape a home — one that can be stored, queried, traversed, and evolved. Together, they transform agent memory from a passive log of past text into an active, organized, self-improving model of the agent’s world.
本體論定義了你的智能體所理解之事的形狀。圖則為那個形狀提供了一個歸宿——一個可被儲存、查詢、遍歷與演化的歸宿。兩者結合，便將智能體的記憶從一份被動的過往文字日誌，轉化為一個主動、有組織、會自我改進的智能體世界模型。
This is the direction we’re pursuing, and it’s encouraging to see it validated by a broad and growing body of research. The agents that remember best won’t be the ones with the most data. They’ll be the ones who understand the structure of what they know.
這正是我們所追尋的方向，而看到它獲得一個廣泛且持續增長的研究體系所驗證，著實令人振奮。記得最好的智能體，不會是擁有最多資料的那些，而會是理解自己所知之事結構的那些。
This article was inspired by “Graph-based Agent Memory: Taxonomy, Techniques, and Applications” (Yang et al., 2026), a comprehensive survey covering 200+ works on memory systems for LLM agents.
本文受到《基於圖的智能體記憶：分類體系、技術與應用》(“Graph-based Agent Memory: Taxonomy, Techniques, and Applications”，Yang 等人，2026) 的啟發，這是一篇全面性的綜述，涵蓋了 200 多篇關於 LLM 智能體記憶系統的著作。
🔤 關鍵術語¶
| 英文 | 繁中譯名 | 文章中的脈絡 / 簡短說明 |
|---|---|---|
| Ontology | 本體論 / 本體 | Agent 世界中事物類型及其關係的形式化定義，是「理解的綱要（schema for understanding）」 |
| Graph-based storage | 圖譜式儲存 | 能忠實表示、查詢並隨時間演化結構的儲存方式，與本體論互補 |
| Knowledge graph | 知識圖譜 | 以實體—關係三元組（entity-relationship triples）儲存資訊，是 agent 的事實骨幹 |
| Entity-relationship triples | 實體—關係三元組 | 如 (Customer, subscribedTo, EnterprisePlan)，由本體論定義其類型與關係 |
| Vector database / vector store | 向量資料庫 | 多數團隊用來嵌入並以相似度檢索的方案，難以可靠長期運作 |
| Vector embeddings / embedding similarity | 向量嵌入 / 嵌入相似度 | 將內容嵌入後以語意相似度檢索，依賴「相似度的偶然性」 |
| Knowledge memory | 知識記憶 | Agent 對世界組織方式的穩定理解，本質上即本體論，提供鷹架 |
| Experience memory | 經驗記憶 | 實際發生事件的動態紀錄，填充知識記憶的鷹架 |
| Hierarchical graph | 階層式圖譜 | 表達本體論的包含與抽象關係（如 Department→Teams→Members），支援不同粒度推理 |
| Temporal graph | 時序圖譜 | 將三元組擴充為四元組（加上 valid_until），追蹤世界演化 |
| Quadruple | 四元組 | 帶時間戳的三元組，如 (Customer, subscribedTo, ProPlan, valid_until: Tuesday) |
| Temporal invalidation | 時序失效 | 透過建立/到期時間戳解決矛盾，而非直接覆寫（如 Graphiti 系統） |
| Hypergraph / hyperedge | 超圖 / 超邊 | 處理跨越兩個以上實體的關係，將多方互動表示為單一超邊 |
| Hybrid architecture | 混合架構 | 結合多種圖譜，分離靜態本體知識與動態經驗資料 |
| Structured traversal | 結構化遍歷 | 沿本體論定義的具型別關係（typed relationships）逐跳導航檢索 |
| Multi-hop reasoning | 多跳推理 | 串接事實的能力，從錨節點沿關係邊向外擴展，向量檢索無法重現 |
| Rule-based filtering | 規則式過濾 | 利用本體論約束施加時間窗、型別限制或關係限制 |
| Policy-based retrieval | 策略式檢索 | 將檢索視為規劃問題，agentic retriever 依學習到的策略導航圖譜 |
| Anchor node | 錨節點 | 語意檢索找到的相關起始節點，再由此展開結構化遍歷 |
| Memory lifecycle | 記憶生命週期 | 四階段：抽取、儲存、檢索、演化（extraction, storage, retrieval, evolution） |
| Associative inference | 關聯推論 | 發掘節點間未明確編碼的潛在連結，等同 agent 探索自身記憶圖譜的「思考」 |