Is Agent Memory a Database? Rethinking Data Foundations for Long-Term AI Agent Memory

# Is Agent Memory a Database? Rethinking Data Foundations for Long-Term AI Agent Memory
Source: [https://arxiv.org/html/2605.26252](https://arxiv.org/html/2605.26252)
###### Abstract\.

Long\-running AI agents need persistent memory\. Memory supports learning across sessions, reduces repeated context injection, and enables auditing of past decisions\. Current agent memory systems and database paradigms treat memory as storage\. They localize correctness at records, embeddings, or edges\. Each supplies only some of the capabilities that long\-term memory requires\. The result is four recurring failure modes: unregulated growth, missing semantic revision, capacity\-driven forgetting, and read\-only retrieval\. In our vision, long\-term agent memory is a new data\-management workload\. Its correctness is a property of the state trajectory, not of individual records\. We formalize this as Governed Evolving Memory \(GEM\)\. GEM replaces record\-level database operations with four state\-level operators: ingestion, revision, forgetting, and retrieval\. Six correctness conditions govern how the state evolves\. Three structural observations establish that no record\-level system can satisfy these conditions, regardless of the storage model\. We realize the abstraction inMemState, a prototype on a property\-graph backend\.MemStatevalidates feasibility and exposes the gap to a native engine\. We outline three research directions that define memory\-centric data management as a workload\.

††copyright:noneKey Contributions•A data\-centric formalization of long\-term agent memory asMt=\(Dt,St,Pt\)M\_\{t\}=\(D\_\{t\},S\_\{t\},P\_\{t\}\)with four operators \(ingestion, revision, forgetting, retrieval\) and six correctness conditions•A reframing of agent memory as a new data\-management workload, identifying where existing data systems fall short•MemState: a prototype built on an embedded property graph engine demonstrating the proposed foundations•A research agenda for governed, evolving memory in long\-lived AI agentsCode:[https://github\.com/CoDS\-GCS/MemState](https://github.com/CoDS-GCS/MemState)

## 1\.Introduction

![Refer to caption](https://arxiv.org/html/2605.26252v1/x1.png)Figure 1\.Agent memory as an append\-only record store over three weekly snapshots\. Built on database operations, it appends new records and evicts old ones by age, never consolidating the state\. This exposes four failures \(①, ②, ③, and ④\)\.Table 1\.Coverage of the four capabilities required by governed memory across database paradigms and agent memory systems\. Each cell describes how the family addresses the capability\. “None” denotes no native support\. No family supports all four, and each contributes one substrate strength the others lack\.FamilyRelevance\-drivenretentionDependency\-awarepropagationGradedattenuationState\-modifyingretrievalSubstrate strengthDatabase paradigmsRelationalNoneForeign key onlyNoneNoneSchema, ACIDKey\-value / DocumentNoneNoneNoneNoneFlexible or no schemaRDF / Property GraphNoneTyped, entity\-grainNoneNoneTyped structural relationsTemporal DBNoneNoneNoneNoneVersioned historiesVector DBNoneGeometric proximityNoneNoneSemantic similarityAgent memory systemsTiered \(MemGPT, MemOS\)NoneNoneNoneNoneTwo\-level pagingFact\-extraction \(Mem0\)NoneNoneNoneNoneAtomic fact maintenanceGraph\-structured \(Zep\)NoneSingle\-edge invalidationNoneNoneBi\-temporal edgesConsolidation\-based \(MIRIX, EverMemOS\)NoneNoneForesight expiryNoneTyped components, scene consolidationRL\-driven \(Mem\-α\\alpha, Memory\-R1\)NoneNoneNoneNoneLearned update policiesGenerative AgentsImportance rankingNoneNoneRecency on accessImportance \+ reflectionAI agents operate as persistent systems that interact with users, tools, and environments\(CrewAI Inc\.,[2025](https://arxiv.org/html/2605.26252#bib.bib15); OpenAI,[2025b](https://arxiv.org/html/2605.26252#bib.bib36); LangChain Inc\.,[2026](https://arxiv.org/html/2605.26252#bib.bib29)\)\. Unlike question answering systems\(Omaret al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib64),[2026](https://arxiv.org/html/2605.26252#bib.bib65)\), they must maintain and revise information across sessions\(Tanet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib55); Huet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib22)\)\. To support this behavior, agents persist information in external memory beyond the context window\(Packeret al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib38); Liet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib25); Xuet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib59); Wang and Chen,[2025](https://arxiv.org/html/2605.26252#bib.bib57); Rasmussenet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib47)\)\. This persistent state shapes whether agent behavior remains stable as interactions accumulate or its performance degrades\(Wuet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib58); Huet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib22)\)\. Long\-term memory therefore changes what agents can do\. It lets agents learn across tasks and sessions by carrying prior decisions and constraints forward\. It also reduces inference cost and latency by avoiding repeated context injection\.

Current memory designs do not preserve these properties\. Most instead follow an accumulation strategy that continuously appends new information while leaving stored entries unchanged\(LangChain Inc\.,[2026](https://arxiv.org/html/2605.26252#bib.bib29); OpenAI,[2025b](https://arxiv.org/html/2605.26252#bib.bib36); Wang and Chen,[2025](https://arxiv.org/html/2605.26252#bib.bib57); Xuet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib59); Rasmussenet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib47)\)\. Figure[1](https://arxiv.org/html/2605.26252#S1.F1)traces this behavior over three weekly snapshots of the same memory\. Each column represents the memory state at one week, and each box represents a stored entry\. From Week 0 to Week 1, new records are added and redundant entries accumulate\. As the state grows, updated facts remain beside obsolete ones\. By Week 2, older entries are evicted by age rather than by their importance to the user\. Retrieval then operates over this evolving clutter\. These behaviors already appear in daily LLM applications\. ChatGPT\(OpenAI,[2025a](https://arxiv.org/html/2605.26252#bib.bib48)\)and Claude\(Anthropic,[2025b](https://arxiv.org/html/2605.26252#bib.bib51)\)retain user preferences yet still surface outdated facts as current\. Cursor\(Anysphere,[2025](https://arxiv.org/html/2605.26252#bib.bib50)\)and Claude Code\(Anthropic,[2025a](https://arxiv.org/html/2605.26252#bib.bib49)\)learn a codebase yet lose earlier decisions as context grows\. The cost falls on the user\. Users re\-explain context the system has already seen\. They pay rising inference cost as context grows\.

These systems inherit record\-level CRUD operations \(create, read, update, delete\) from traditional databases\. As a result, memory operations act on individual records rather than on the evolving memory state itself\. This mismatch produces the four recurring failure modes shown in Figure[1](https://arxiv.org/html/2605.26252#S1.F1)\.

①*Unregulated growth\.*Append\-only ingestion accumulates redundant and low\-relevance entries\. When the user re\-explains a task, the same facts are ingested again \(e\.g\., the project record“Website Redesign\|\|Deadline: March 15”is stored twice at Week 1\)\. These redundant entries compete at retrieval time and consume LLM context window space, crowding out useful content\.

②*Missing semantic revision\.*Updates are appended rather than integrated into existing entries\. At Week 1,“Deadline UPDATED: April 20”is stored as a new entry while “Deadline: March 15” remains\. A query “What is the deadline of Website Redesign?” may return “March 15” instead of “April 20” \(based on semantic similarity between the query and message embeddings\)\(Tanet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib55); Huet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib22)\)\.

③*Absence of selective forgetting\.*Memory must evict content as storage fills\. But eviction is driven by age or capacity rather than by importance to the user\. At Week 2 in Figure[1](https://arxiv.org/html/2605.26252#S1.F1), the project deadline is evicted while the low\-relevance entry“Discussed lunch preferences”persists\. The same deadline query now returns “I do not know,” even though the user asked it before\. The system cannot retain facts according to their relevance to the user\(Maharanaet al\.,[2024](https://arxiv.org/html/2605.26252#bib.bib26)\)\.

④*Read\-only retrieval\.*Retrieval returns facts but never updates the memory state\(Wuet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib58)\)\. The user queries the project deadline every week, yet that entry gains no importance and is later evicted like any other \(Failure ③\)\. User interaction patterns cannot reinforce useful information or protect it from forgetting\. Frequently accessed facts therefore compete with stale content on equal terms\.

These limitations reflect an abstraction gap, not implementation issues\. Each failure traces to one CRUD operation: create cannot integrate, update cannot propagate, delete cannot regulate relevance, and read cannot adapt\. Larger context windows or better retrieval do not resolve this mismatch\(Tanet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib55); Huet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib22)\)\. The limitation lies in missing evolution semantics, not in retrieval quality\.

Contributions\.This paper positions long\-term agent memory as a new data\-management workload whose correctness lives in the state trajectory, not in individual records\. Our contributions are:

- •Afour\-capability analytical lens\(relevance\-driven retention, dependency\-aware propagation, graded attenuation, and state\-modifying retrieval\) showing that no database paradigm or agent memory system supplies all four \(Section[2](https://arxiv.org/html/2605.26252#S2)\)\.
- •Governed Evolving Memory \(GEM\), a state abstraction that replaces record\-level CRUD with four state\-level operators \(ingestion, revision, forgetting, retrieval\) and defines six correctness conditions over the state trajectory\. Three structural observations show that no CRUD\-based system can satisfy these conditions, regardless of substrate \(Section[3](https://arxiv.org/html/2605.26252#S3)\)\.
- •MemState, a prototype that realizes GEM on a property\-graph backend with topic\-based storage, typed dependencies, and declarative policies\. The prototype validates feasibility and exposes what a native engine must provide \(Section[4](https://arxiv.org/html/2605.26252#S4)\)\.
- •Aresearch agendaof three directions covering a native engine, trajectory\-level correctness, and privacy under multi\-tenant memory, with explicit success criteria\. \(Section[5](https://arxiv.org/html/2605.26252#S5)\)\.

## 2\.Why Current Abstractions Fail

We examine database paradigms and recent agent memory systems through four capabilities required by governed memory\. Each is a column in Table[1](https://arxiv.org/html/2605.26252#S1.T1)and a failure mode in Figure[1](https://arxiv.org/html/2605.26252#S1.F1)\. Table[1](https://arxiv.org/html/2605.26252#S1.T1)covers five database paradigms and five families of agent memory systems, plus Generative Agents\(Parket al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib8)\)as the closest single approach\. Each contributes one substrate strength\. None covers all four capabilities\.

Database paradigms and memory families\.Database paradigms differ by what they store and how they update\. Relational stores manage records under a fixed schema\(Codd,[1970](https://arxiv.org/html/2605.26252#bib.bib9)\)\. Key\-value and document stores relax this, using no schema or a flexible one\(DeCandiaet al\.,[2007](https://arxiv.org/html/2605.26252#bib.bib16); Changet al\.,[2008](https://arxiv.org/html/2605.26252#bib.bib13)\)\. RDF and property graphs add typed relations between entities\(Angles and Gutierrez,[2008](https://arxiv.org/html/2605.26252#bib.bib11); Franciset al\.,[2018](https://arxiv.org/html/2605.26252#bib.bib19); Neumann and Weikum,[2010](https://arxiv.org/html/2605.26252#bib.bib34); Perezet al\.,[2009](https://arxiv.org/html/2605.26252#bib.bib46)\)\. Temporal databases version tuples to preserve history\(Jensen and Snodgrass,[2002](https://arxiv.org/html/2605.26252#bib.bib24); Snodgrass,[1999](https://arxiv.org/html/2605.26252#bib.bib62)\)\. Vector databases index embeddings for semantic similarity\(Wanget al\.,[2021](https://arxiv.org/html/2605.26252#bib.bib61); Panet al\.,[2024a](https://arxiv.org/html/2605.26252#bib.bib44)\)\.

Agent memory systems group into five families by their primary mechanism\. Tiered designs \(MemGPT\(Packeret al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib38)\), MemOS\(Liet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib25)\)\) implement two\-level paging: a small in\-context active tier and a larger external storage tier\. Evicting from the active tier by age or size when it fills\. Fact\-extraction systems \(Mem0\(Chhikaraet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib14)\)\) parse interactions into atomic facts and overwrite on conflict\. Graph\-structured systems \(Zep\(Rasmussenet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib47)\)\) link entries via typed edges and invalidate them bi\-temporally to preserve history\. Consolidation\-based systems \(MIRIX\(Wang and Chen,[2025](https://arxiv.org/html/2605.26252#bib.bib57)\), EverMemOS\(Huet al\.,[2026](https://arxiv.org/html/2605.26252#bib.bib71)\)\) route content into specialized memory types and cluster related entries into higher\-level structures\. RL\-driven systems \(Mem\-α\\alpha\(Wanget al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib73)\), Memory\-R1\(Yanet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib72)\)\) learn update policies by reinforcement, rewarding operations by downstream answer quality\. Generative Agents\(Parket al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib8)\)rank memories by importance and update recency on read\. Each family contributes one substrate strength; Table[1](https://arxiv.org/html/2605.26252#S1.T1)shows where each falls short\. The rest of this section examines each capability in turn\.

### 2\.1\.Relevance\-Driven Retention

Relevance\-driven retention bounds active memory by utility, stabilizing inference cost as interactions grow\. Without it, memory grows monotonically and redundant entries crowd out useful ones \(failure mode ①\)\.

Database paradigmscap growth by capacity or time, not utility\. Relational and document stores use TTL expiry or manual delete\(Elmasri and Navathe,[2016](https://arxiv.org/html/2605.26252#bib.bib18)\); temporal databases enforce retention windows on versioned tuples\(Jensen and Snodgrass,[2002](https://arxiv.org/html/2605.26252#bib.bib24)\); vector databases prune embeddings by capacity or age\(Wanget al\.,[2021](https://arxiv.org/html/2605.26252#bib.bib61); Panet al\.,[2024b](https://arxiv.org/html/2605.26252#bib.bib45)\)\. High\- and low\-utility facts age out equally \(in Figure[1](https://arxiv.org/html/2605.26252#S1.F1),“Discussed lunch preferences”expires on the same schedule as the project deadline\)\. No paradigm bounds memory by relevance\.

Agent memory systemsrepeat the pattern\. Tier\-based, lifecycle\-based, capacity\-based, and learned\-controller eviction all key on age or size\(Packeret al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib38); Liet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib25); Huet al\.,[2026](https://arxiv.org/html/2605.26252#bib.bib71); Chhikaraet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib14); Wanget al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib73); Yanet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib72)\)\. Generative Agents\(Parket al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib8)\)come closest\. They combine an importance score with recency decay to rank observations at retrieval\. Both are local heuristics applied at retrieval time, not policies that bound the active footprint\. Memory grows because low\-importance observations are never attenuated\.

### 2\.2\.Dependency\-Aware Propagation

Dependency\-aware propagation keeps related facts consistent when one changes\. The agent then reasons over coherent state rather than divergent values\. Without it, contradictions yield wrong answers \(failure mode ②\)\.

Database paradigmsoffer two relevant techniques, neither propagating along semantic dependencies\. Active databases\(Widom and Ceri,[1996](https://arxiv.org/html/2605.26252#bib.bib1); Ceri and Widom,[1990](https://arxiv.org/html/2605.26252#bib.bib3)\)propagate via ECA rules over flat relational state, so updates are referential, not content\-aware\. Materialized view maintenance\(Olteanu,[2024](https://arxiv.org/html/2605.26252#bib.bib2)\)propagates along fixed view schemas, not evolving semantic units\. Elsewhere, append\-only relational and vector stores return outdated and current values\(Codd,[1970](https://arxiv.org/html/2605.26252#bib.bib9); Wanget al\.,[2021](https://arxiv.org/html/2605.26252#bib.bib61); Panet al\.,[2024a](https://arxiv.org/html/2605.26252#bib.bib44)\); in\-place updates destroy the evidence chain\(DeCandiaet al\.,[2007](https://arxiv.org/html/2605.26252#bib.bib16); Changet al\.,[2008](https://arxiv.org/html/2605.26252#bib.bib13)\); temporal stores return superseded values as current\(Jensen and Snodgrass,[2002](https://arxiv.org/html/2605.26252#bib.bib24); Snodgrass,[1999](https://arxiv.org/html/2605.26252#bib.bib62)\); and property graph or RDF stores desynchronize at entity grain\(Angles and Gutierrez,[2008](https://arxiv.org/html/2605.26252#bib.bib11); Franciset al\.,[2018](https://arxiv.org/html/2605.26252#bib.bib19); Fenget al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib41); Neumann and Weikum,[2010](https://arxiv.org/html/2605.26252#bib.bib34); Perezet al\.,[2009](https://arxiv.org/html/2605.26252#bib.bib46)\)\. No DBMS re\-evaluates dependent facts when one fact changes\.

Agent memory systemsinherit the gap\. Mem0\(Chhikaraet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib14)\)overwrites or deletes on conflict, destroying the evidence chain\. Zep\(Rasmussenet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib47)\)invalidates edges bi\-temporally\. It is the strongest mechanism but operates one edge at a time\. In Figure[1](https://arxiv.org/html/2605.26252#S1.F1), invalidating the deadline edge at Week1 does not re\-evaluate edges to team assignments or meetings\. Dependencies then drift silently\. Consolidation\-based systems\(Wang and Chen,[2025](https://arxiv.org/html/2605.26252#bib.bib57); Huet al\.,[2026](https://arxiv.org/html/2605.26252#bib.bib71)\)overwrite or re\-cluster at component grain, not along dependencies\. RL\-driven systems\(Wanget al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib73); Yanet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib72)\)update without dependency structure\.

### 2\.3\.Graded Attenuation

Graded attenuation deprioritizes obsolete content while preserving history for audit, unlike tiering, which moves whole entries by capacity\. Without it, obsolete entries compete with current ones \(failure mode ③\)\.

Database paradigmsattenuate by time or capacity\. TTL, retention windows, and version pruning act on age and size, not importance\(Jensen and Snodgrass,[2002](https://arxiv.org/html/2605.26252#bib.bib24); Snodgrass,[1999](https://arxiv.org/html/2605.26252#bib.bib62); Wanget al\.,[2021](https://arxiv.org/html/2605.26252#bib.bib61)\)\. Removal is binary\. No paradigm demotes content while keeping it recoverable\. In Figure[1](https://arxiv.org/html/2605.26252#S1.F1), the high\-utility project deadline ages out while a low\-relevance entry survives\.

Agent memory systemsevict by FIFO, lifecycle stage, or store size\(Packeret al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib38); Liet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib25); Huet al\.,[2026](https://arxiv.org/html/2605.26252#bib.bib71)\), or delete locally\(Chhikaraet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib14)\), without relevance\. EverMemOS\(Huet al\.,[2026](https://arxiv.org/html/2605.26252#bib.bib71)\)is the partial exception\. It expires time\-bounded foresight at retrieval while keeping the rest\. Zep\(Rasmussenet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib47)\)caches ingestion\-time summaries that aid retrieval but can inflate the store\(Chhikaraet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib14)\)\. Generative Agents\(Parket al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib8)\)rank observations at retrieval by recency, importance, and relevance but never attenuate them from the store\.

### 2\.4\.State\-Modifying Retrieval

State\-modifying retrieval updates salience on each read so important content stays prominent and stale content fades\. Without it, accessed and stale facts compete equally \(failure mode ④\)\.

Database paradigmstreat retrieval as a pure read\. A query returns content and leaves state unchanged\. Recent retrieval indexes scale read\-only access\(Malkov and Yashunin,[2018](https://arxiv.org/html/2605.26252#bib.bib32)\)but do not let a read update the state it reads\.

Agent memory systemsdo the same\. In Figure[1](https://arxiv.org/html/2605.26252#S1.F1), the Week1 deadline query returns the stored value without reinforcing it, propagating access, or shaping future retrievals\. Zep\(Rasmussenet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib47)\)reranks by mention frequency, a read\-time heuristic, not a committed state change\. Generative Agents\(Parket al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib8)\)update a memory’s recency on access, but only as an embedded heuristic with no correctness condition\. Governed memory needs retrieval as a first\-class operator that returns output and a state transition \(Section[3](https://arxiv.org/html/2605.26252#S3)\)\.

⋆\\starIs Agent Memory a Database?No\. CRUD governs records; correctness here lives in the state trajectory\. Agent memory is a new data\-management workload, not a database problem\.

## 3\.Governed Evolving Memory

This section introduces our Governed Evolving Memory \(GEM\), the abstraction that supplies the four capabilities\. Memory is a global state that evolves through structured operations rather than append\-only accumulation\. Figure[2](https://arxiv.org/html/2605.26252#S3.F2)illustrates GEM\.

![Refer to caption](https://arxiv.org/html/2605.26252v1/x2.png)Figure 2\.Our GEM Abstraction\. The stateMt=\(Dt,St,Pt\)M\_\{t\}=\(D\_\{t\},S\_\{t\},P\_\{t\}\)holds semantic units \(DtD\_\{t\}\), their structural organization \(StS\_\{t\}\), and declarative evolution policies \(PtP\_\{t\}\)\. These three elements must be explicit in any compliant implementation; their realization varies by backend\. Four state\-level operators replace record\-level CRUD: ingestion, revision, forgetting, retrieval\.### 3\.1\.Memory State

Three elements must be explicit in the data model\. ContentDtD\_\{t\}specifies what is stored\. StructureStS\_\{t\}specifies how stored elements connect\. PoliciesPtP\_\{t\}specify how the state is allowed to change\. If any element remains implicit, evolution semantics cannot be enforced\. A*semantic unit*is the atom ofDtD\_\{t\}, carrying a value history and a salience signal\. Its boundary is a design decision\. Finer units scatter related facts across boundaries, requiring graph traversal at retrieval and risking partial recall\.

###### Definition 0\(Governed Evolving State\)\.

At timett, an agent’s memory state is the tupleMt=\(Dt,St,Pt\)M\_\{t\}=\(D\_\{t\},S\_\{t\},P\_\{t\}\), whereDtD\_\{t\}denotes stored content organized as semantic units that hold all related data elements,StS\_\{t\}denotes the structural organization over that content, andPtP\_\{t\}denotes policies governing access, ingestion, revision, and forgetting\.

###### Definition 0\(Memory Evolution\)\.

Memory evolves according to a state transition functionMt\+1=𝒰​\(Mt,It,Rt\)M\_\{t\+1\}=\\mathcal\{U\}\(M\_\{t\},I\_\{t\},R\_\{t\}\), whereItI\_\{t\}is new external input andRtR\_\{t\}is an internal operation that may alter state\. The sequence\{Mt\}t≥0\\\{M\_\{t\}\\\}\_\{t\\geq 0\}is the memory trajectory\.

###### Definition 0\(Evolution Policies\)\.

PtP\_\{t\}is a set of typed rules⟨\\langle*event*,*condition*,*action*⟩\\rangle, where*event*identifies the operation that triggers evaluation,*condition*is a predicate overMtM\_\{t\}, and*action*is a state\-level transition\. Policies are declarative: they specify what transitions occur and when, independent of how operators execute them\.

### 3\.2\.State\-Level Operators

Four operators act over global memory under policy constraints, each supplying one capability from Section[2](https://arxiv.org/html/2605.26252#S2)\.

Ingestionintegrates inputItI\_\{t\}into the existing state by producingMt\+1=𝒰​\(Mt,It,∅\)M\_\{t\+1\}=\\mathcal\{U\}\(M\_\{t\},I\_\{t\},\\emptyset\)under constraints inPtP\_\{t\}\. An updated value is recorded against the existing semantic unit; the prior value is retained as historical evidence\.

RevisionproducesMt\+1=𝒰​\(Mt,∅,rev​\(Δ\)\)M\_\{t\+1\}=\\mathcal\{U\}\(M\_\{t\},\\emptyset,\\textsc\{rev\}\(\\Delta\)\)from internal evidenceΔ\\DeltaunderStS\_\{t\}andPtP\_\{t\}\. It reconciles overlapping units, propagates updates alongStS\_\{t\}, and preserves superseded values with provenance\.

Forgettingis a policy\-governed transitionMt\+1=𝒰​\(Mt,∅,ℱ\)M\_\{t\+1\}=\\mathcal\{U\}\(M\_\{t\},\\emptyset,\\mathcal\{F\}\), whereℱ\\mathcal\{F\}regulates the influence of stored content by relevance signals without destructive deletion\. Content inDtD\_\{t\}carries salience signals that rise on access and decay on disuse, at sub\-unit granularity, so part of a unit may be attenuated while the rest stays current\. Attenuation runs as a ladder from partial compression to full archiving, so the operator is graded rather than binary\.

Retrievalmaps queryqqto outputooviaℛ​\(Mt,q\)→o\\mathcal\{R\}\(M\_\{t\},q\)\\to oand inducesMt\+1=𝒰​\(Mt,∅,ℛq\)M\_\{t\+1\}=\\mathcal\{U\}\(M\_\{t\},\\emptyset,\\mathcal\{R\}\_\{q\}\)underPtP\_\{t\}\. Every read updates the salience of accessed units, so retrieval is a state transition rather than a read\-only operation\.

### 3\.3\.Memory Correctness

Correctness cannot be defined at the level of individual records, because contradictory or outdated entries may coexist withinDtD\_\{t\}\. Correctness is a property of the trajectory along three axes: what queries return, what the state preserves, and how the state adapts\.

###### Definition 0\(Memory Correctness\)\.

Letui∈Dtu\_\{i\}\\in D\_\{t\}be a semantic unit with a value historyVi=⟨\(v1,t1,π1\),…,\(vk,tk,πk\)⟩V\_\{i\}=\\langle\(v\_\{1\},t\_\{1\},\\pi\_\{1\}\),\\ldots,\(v\_\{k\},t\_\{k\},\\pi\_\{k\}\)\\rangle, where each entry records a valuevv, a timestamptt, and provenanceπ\\pi\. A memory system is*correct*if the following six conditions hold\.

C1 \(Query soundness\)\.The responseℛ​\(Mt,q\)→o\\mathcal\{R\}\(M\_\{t\},q\)\\to oreflects the most recent non\-archived valuevkv\_\{k\}as current; prior values appear only whenqqexplicitly requests historical context\.

C2 \(Transition soundness\)\.Every transitionMt→Mt\+1M\_\{t\}\\to M\_\{t\+1\}respectsPtP\_\{t\}, and no revision produces a state in which a superseded value is returned as current\.

C3 \(Dependency consistency\)\.For every pair\(ui,uj\)\(u\_\{i\},u\_\{j\}\)connected by a typed edgee∈Ste\\in S\_\{t\}with propagation semantics, an update touiu\_\{i\}triggers evaluation ofuju\_\{j\}underPtP\_\{t\}\.

C4 \(Provenance preservation\)\.Forgetting and revision preserve the provenance chain of any unit that remains reachable\.

C5 \(Bounded active state\)\.For every interaction countnn, the active memory satisfies\|Dtactive\|≤β​\(n\)\|D\_\{t\}^\{\\text\{active\}\}\|\\leq\\beta\(n\)for a policy\-defined boundβ\\beta; archived content remains recoverable\.

C6 \(Retrieval\-induced adaptation\)\.Every retrieval that accessesuiu\_\{i\}induces a transition in which the salience ofuiu\_\{i\}is updated\. Repeated retrieval strictly reducesuiu\_\{i\}’s eligibility for attenuation\.

C1–C2 govern what queries return\. C3–C4 govern what the state preserves\. C5–C6 govern how the state adapts\. They define correctness as a property of\{Mt\}t≥0\\\{M\_\{t\}\\\}\_\{t\\geq 0\}, not of any individual record\. Every failure mode in Figure[1](https://arxiv.org/html/2605.26252#S1.F1)reduces to the violation of at least one\.

### 3\.4\.Three Structural Observations

The four operators are not a recombination of CRUD:

Observation 1 \(Retrieval\)\.A pure\-function retrieval operator cannot satisfy C6\. C6 requires retrieval itself to induce a state transition\. Caches, materialized views, and post\-retrieval triggers can record that an access occurred, but they cannot lift the operator out of being a pure function, because the state\-modifying step is decoupled from the query\.

Observation 2 \(Forgetting\)\.C5 in its relevance\-driven form is jointly unenforceable with C6 above any CRUD\-based engine\. If retrieval is read\-only, relevance must be approximated by external signals, and capacity\-driven or time\-based attenuation can bound size but not relevance\.

Observation 3a \(Ingestion\)\.Append\-only storage without semantic units cannot satisfy C2\. Two appended values for the same fact coexist with equal status, and a default query has no engine\-level mechanism to select between them\.

Observation 3b \(Revision\)\.Untyped propagation cannot satisfy C3\. Updates propagate along exact\-match references or untyped edges, so dependencies that the abstraction requires to re\-evaluate are not visible to the engine\.

These are structural claims, not theorems\. Their consequence: governed memory requires four state\-level operators inside the data model, withPtP\_\{t\}checked at commit and retrieval treated as a write\.

## 4\.Realizing GEM in MemState

We instantiate GEM \(Section[3](https://arxiv.org/html/2605.26252#S3)\) in our prototype,MemState111Code:[https://github\.com/CoDS\-GCS/MemState](https://github.com/CoDS-GCS/MemState)\.\. The data model supplies C1, C3, and C4 by construction and lifts C2 to a data\-model guarantee\. The four operators of Section[3\.2](https://arxiv.org/html/2605.26252#S3.SS2)run over this data model and supply C5 and C6 under transitions\.

### 4\.1\.Data Model

MemState realizesMt=\(Dt,St,Pt\)M\_\{t\}=\(D\_\{t\},S\_\{t\},P\_\{t\}\)as an evolving graph of*topics*on Kuzu\(Fenget al\.,[2023](https://arxiv.org/html/2605.26252#bib.bib41)\), an embedded property graph engine\. Each topic is a self\-contained semantic unit with a title, a summary, a dense embedding, and a set of fields with value histories \(Figure[3](https://arxiv.org/html/2605.26252#S4.F3)\(a\)\)\.

Self\-contained topics with field histories \(C1, C4\)\.A topic groups all fields of one concept in a single unit; entity\-grain designs \(Zep\) scatter attributes across nodes, requiring multiple accesses to reconstruct one concept\. Each field is maintained as a historyHi,j=⟨\(v,t,π\)⟩H\_\{i,j\}=\\langle\(v,t,\\pi\)\\rangle: updates append a new entry rather than overwrite\. Default retrieval returns onlyvkv\_\{k\}\(C1\); explicit temporal queries can read anyvjv\_\{j\}and its provenance \(C4\)\. Once sufficient knowledge accumulates around a subset of fields, revision promotes it into a standalone topic:*Alice*splits out of Website Redesign once enough interactions reference her directly \(Figure[3](https://arxiv.org/html/2605.26252#S4.F3)\(a,b\)\)\.

Typed edges inStS\_\{t\}\(C3\)\.StS\_\{t\}distinguishes two edge types \(Figure[3](https://arxiv.org/html/2605.26252#S4.F3)\(b\)\)\.*Extension*edges connect topics where a change in one can entail a change in the other; a deadline change on*Website Redesign*may entail a milestone reschedule\.*Association*edges connect related but independent topics\. C3 propagation must follow entailment, not relatedness, so revision traverses extension edges only\. Association edges support retrieval context expansion without propagation\.

![Refer to caption](https://arxiv.org/html/2605.26252v1/x3.png)Figure 3\.MemState data model\. \(a\) A self\-contained topic stores fields, values, histories, and provenance; low\-salience content can be forgotten\. \(b\) Topics form a typed graph through*association*and*extension*edges; revision may promote subsets \(e\.g\.,*Alice*\) into new topics\.Declarative policiesPtP\_\{t\}in state \(C2\)\.Policies live insideMtM\_\{t\}as⟨\\langle*event*,*condition*,*action*⟩\\ranglerules whose conditions referenceMtM\_\{t\}directly \(Definition[3\.3](https://arxiv.org/html/2605.26252#S3.Thmtheorem3)\)\. The postcondition on a proposedMt\+1M\_\{t\+1\}is evaluated againstPtP\_\{t\}before commit, and a violating transition is rejected\. This lifts C2 to a data\-model\-level guarantee, closing Observation 3a\. Listing[1](https://arxiv.org/html/2605.26252#LST1)shows a representative policy\. Policies can be added, modified, or replaced without changing operator code\.

### 4\.2\.Operators on the Kuzu Substrate

State evolution is realized through the four operators of Section[3\.2](https://arxiv.org/html/2605.26252#S3.SS2)\. Ingestion and retrieval are client\-facing and synchronous\. Revision and forgetting run asynchronously as policy\-triggered maintenance\. Algorithm[1](https://arxiv.org/html/2605.26252#alg1)shows their unified template on Kuzu\. An incoming event is dispatched to one of four operator branches \(lines 3–12\), each writing through field\-level, topic\-level, and graph\-level primitives on the property graph\.PtP\_\{t\}is then evaluated against the proposedMt\+1M\_\{t\+1\}and the transition either commits atomically or aborts \(line 14\)\. The atomic commit is the mechanism that lifts C2 to a data\-model\-level guarantee\. The salience increment inside the retrieval branch \(line 11\) is the mechanism for C6\.

Listing 1:A representative MemState policy: when a field changes, mark dependent topics for revision \(C3\)\.POLICYpropagate\-on\-change

ONfield\_updated

WHENEXISTSdependent\_topic

DOflag\_for\_revision\(dependent\_topic\)

WITHevidence=\{updated\_field,timestamp\}

Algorithm 1GEM transition on a property\-graph substrate\.1:Memory state

Mt=\(Dt,St,Pt\)M\_\{t\}=\(D\_\{t\},S\_\{t\},P\_\{t\}\); event

e∈\{It,q,Δ\}e\\in\\\{I\_\{t\},q,\\Delta\\\}
2:Updated state

Mt\+1M\_\{t\+1\}; output

ooif

e=qe=q
3:

𝑜𝑝←𝑑𝑖𝑠𝑝𝑎𝑡𝑐ℎ​\(e\)\\mathit\{op\}\\leftarrow\\mathit\{dispatch\}\(e\)⊳\\trianglerightingest, revise, forget, retrieve

4:begin transaction

5:if

𝑜𝑝=𝑖𝑛𝑔𝑒𝑠𝑡\\mathit\{op\}=\\mathit\{ingest\}then

6:LLM picks host topic

τ\\tau; for each fact

\(f,v,t,π\)\(f,v,t,\\pi\), append to

Hτ,fH\_\{\\tau,f\}or create

ff; refresh

τ\\tau’s embedding; flag extension\-linked topics

7:elseif

𝑜𝑝=𝑟𝑒𝑣𝑖𝑠𝑒\\mathit\{op\}=\\mathit\{revise\}then

8:apply repair \(conflict, merge, propagate\) for each evidence

δ∈Δ\\delta\\in\\Delta
9:elseif

𝑜𝑝=𝑓𝑜𝑟𝑔𝑒𝑡\\mathit\{op\}=\\mathit\{forget\}then

10:attenuate each

uuwith

salience​\(u\)<θ∗\\mathrm\{salience\}\(u\)<\\theta\_\{\*\}\(compress, hide, archive\)

11:elseif

𝑜𝑝=𝑟𝑒𝑡𝑟𝑖𝑒𝑣𝑒\\mathit\{op\}=\\mathit\{retrieve\}then

12:route

qq; read selected units; build

oo; increment salience of accessed units⊳\\trianglerightC6

13:endif

14:evaluate

PtP\_\{t\}on proposed

Mt\+1M\_\{t\+1\};commitif all postconditions hold, elseabort⊳\\trianglerightatomic commit lifts C2

15:return

Mt\+1M\_\{t\+1\}\(and

ooif applicable\)

Ingestion\.The LLM reads topic titles and summaries to select a host topic, then reads the topic schema to place the new value in an existing or new field\. For the deadline update from Figure[1](https://arxiv.org/html/2605.26252#S1.F1), the new entry\(April 20,T1,π1\)\(\\textit\{April 20\},T\_\{1\},\\pi\_\{1\}\)is appended toHdeadlineH\_\{\\textit\{deadline\}\}while\(March 15,T0,π0\)\(\\textit\{March 15\},T\_\{0\},\\pi\_\{0\}\)remains\. The embedding is refreshed and extension\-linked topics are flagged for revision\. C1, C2, and C4 hold by construction\.

Revision\.Revision detects evidence items inMtM\_\{t\}\(duplicate topics, conflicting field values, schema drift, dependency inconsistencies\) and applies the corresponding repair\. Dependency repair walks extension edges, halting at any topic whose policy condition does not fire\. Topic granularity keeps this frontier far smaller than an entity\-grain graph \(Zep\), so the walk terminates in few hops\. This supplies the dependency\-aware propagation capability of Section[2\.2](https://arxiv.org/html/2605.26252#S2.SS2)at the global level\. Conflict resolution marks superseded values with their provenance rather than deleting them, preserving C4\.

Forgetting\.Each field maintains a salience score that rises on access and decays on disuse\. Three thresholds define a graded ladder\. Belowθ𝑠𝑢𝑚𝑚𝑎𝑟𝑦\\theta\_\{\\mathit\{summary\}\}, a history is compressed\. Belowθ𝑟𝑒𝑚𝑜𝑣𝑒\\theta\_\{\\mathit\{remove\}\}, a field is hidden from active retrieval\. Belowθ𝑎𝑟𝑐ℎ𝑖𝑣𝑒\\theta\_\{\\mathit\{archive\}\}, the topic is archived but remains recoverable through explicit lookup\. Retention is driven by salience rather than age or capacity, satisfying C5\.

Retrievalroutes a query into topic\-based, temporal, or structural mode\. In topic\-based mode, it inspects titles and summaries, selects candidate topics, reads schemas, and reads required field values\. For*“What is the deadline for the Website Redesign?”*it returns the current*Deadline*value\. The salience increment on every accessed unit is part of the operator semantics, closing Observation 1\.

### 4\.3\.From Prototype to Native Engine

What MemState validates\.MemState is a feasibility sketch on a property\-graph substrate\. GEM is realizable on commodity infrastructure: topics, field histories, embeddings, and policy postcondition checks all attach to one Kuzu transaction, so C1, C2, and C4 hold by construction and C5, C6 hold under transitions\.

What MemState exposes\.The substrate is a compatibility layer, not a native expression of GEM\. Field histories, propagation\-bearing edges, and postcondition\-checked commits are reconstructed from generic graph primitives, so several optimizations are not directly expressible\. A native engine would store field histories as first\-class bitemporal attributes, attach propagation semantics to extension edges in the schema, and compile policy postconditions into the commit protocol\. Retrieval\-induced salience updates would compile into a single read\-modify\-write primitive, and relevance\-driven forgetting would be scheduled like index maintenance\. These are not implementation gaps; they are research directions for a data\-management workload that no current engine targets\.

## 5\.Research Agenda

This section presents three research directions and the success criteria of the vision\. The directions follow from GEM and MemState, grounded in capabilities no current engine supports natively\.

A Native Engine for Governed Memory\.MemState reconstructs topic records, field histories, propagation\-bearing edges, and policy\-checked commits from generic property\-graph primitives, at a compatibility\-layer cost\. A native engine must address three problems\. \(i\)*Storage layout:*co\-locate topics, field histories, and embeddings on pages to reduce I/O cost per read, following the node and neighbor co\-location used in graph\-based vector engines\(Sunet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib39)\), adapted to units that carry both a value history and typed dependencies\. \(ii\)*Unified indexing:*jointly support semantic similarity and history predicates, so vector\-based, temporal, and structural queries route through the same physical organization without duplicating data\. \(iii\)*Retrieval as a write:*C6 requires every read to update salience, but existing query languages separate reads from writes\(Franciset al\.,[2018](https://arxiv.org/html/2605.26252#bib.bib19); Perezet al\.,[2009](https://arxiv.org/html/2605.26252#bib.bib46); Kondylakiset al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib42)\)and existing indexes optimize read\-only access at scale\(Malkov and Yashunin,[2018](https://arxiv.org/html/2605.26252#bib.bib32)\)\. A native engine needs an operator that unifies search, traversal, temporal lookup, and salience updates, with a buffer strategy that keeps hot topics and dependency context resident\. First targets: \(i\) an I/O\-aware page layout for topics with field histories and extension edges, \(ii\) a joint index over semantic similarity and history predicates, and \(iii\) the consistency cost of retrieval\-induced salience updates under concurrent access\.

Correctness and Evaluation\.Correctness in GEM is a property of the state trajectory\{Mt\}t≥0\\\{M\_\{t\}\\\}\_\{t\\geq 0\}, not of individual records\. Three sub\-problems follow\. \(i\)*Trajectory benchmark\.*Current benchmarks measure answer\-level recall and exercise C1 only partially\(Maharanaet al\.,[2024](https://arxiv.org/html/2605.26252#bib.bib26); Wuet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib58); Tanet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib55); Huet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib22)\)\. A system that overwrites history or never forgets can still score well if its recent answers are correct\. A trajectory benchmark needs ground truth at three levels: the current value of each unit over time \(C2\), the dependent units that change after each update \(C3\), and the active footprint at each interaction count \(C5\)\. No existing benchmark provides all three\. \(ii\)*Policy language for conflict resolution\.*Existing systems either overwrite on conflict\(Chhikaraet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib14)\)or invalidate edges through an LLM gate that silently misses contradictions\(Rasmussenet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib47)\)\. A declarative policy language must express conflict\-resolution rules over field histories with provenance, exploiting typed edges rather than LLM gating\. \(iii\)*Constrained learned controllers\.*Mem\-α\\alpha\(Wanget al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib73)\)and Memory\-R1\(Yanet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib72)\)learn update policies over flat fact stores withoutStS\_\{t\}or field histories\. The open problem is a controller that optimizes within trajectory\-level constraints enforced by the commit protocol\. First targets: \(i\) a 500\-turn adversarial workload scoring Mem0, Zep, and MemState on answer\- and trajectory\-level metrics, \(ii\) a policy language prototype on LongMemEval\(Wuet al\.,[2025](https://arxiv.org/html/2605.26252#bib.bib58)\)and LoCoMo\(Maharanaet al\.,[2024](https://arxiv.org/html/2605.26252#bib.bib26)\), and \(iii\) an RL controller trained against C2–C5 constraints\.

Privacy and Multi\-Tenancy\.A third direction concerns shared memory\. Production agents often serve multiple tenants over a common memory instance, which introduces two problems that do not arise in single\-tenant settings\. \(i\)*Retrieval\-induced information leakage\.*C6 makes retrieval a write operation\. If tenantAA’s query reinforces topicτ\\tau, the salience update is committed toMtM\_\{t\}as a state transition\. A later query from tenantBBsurfacesτ\\tauthrough similarity ranking because its salience score is high\. The salience signal acts as an information leakage path across tenant isolation boundaries, even when topic content is access\-controlled\. Existing memory systems treat retrieval as a read\-only operation and do not account for this side effect\. \(ii\)*Verifiable erasure under evolving state\.*Privacy regulations require provable removal of a tenant’s data upon request\. GEM makes erasure strictly harder than relational delete, because a forgetting operator must compose with C4 \(provenance preservation\) and C6 \(derived salience\)\. A tenant’s data shapes provenance chains on other topics and salience aggregates that influenced other tenants’ query results\. Deleting the base records does not erase these derived signals\. Data exchange research\(Faginet al\.,[2005](https://arxiv.org/html/2605.26252#bib.bib43)\)addresses cross\-boundary consistency but not privacy\-preserving forgetting over an evolving state trajectory\. First targets: \(i\) the information leakage rate between two tenants on a shared MemState instance and a retrieval operator with bounded salience side effects, and \(ii\) a forgetting operator that erases derived salience and provenance traces under C4 and C6\.

Success Criteria\.The vision succeeds when four conditions hold\. \(i\) At least one DBMS exposes governed\-evolution operators as first\-class primitives, and agent frameworks declare retention and propagation policies declaratively\. \(ii\) Standardized trajectory\-level benchmarks measure C2–C6 violations across long interaction histories\. \(iii\) Long\-horizon deployments show measurable reductions in temporal\-reasoning errors attributable to GEM\-conformant memory\. \(iv\) Privacy and forgetting guarantees over evolving memory become as well\-understood as ACID guarantees over transactional storage\. The arc parallels stream processing, which became a recognized workload once continuous state and event\-time semantics moved from application code into the data model\.

## 6\.Conclusion

Long\-term agent memory is the workload behind every persistent AI agent, but no current system treats it as one\. We argued that its correctness is a property of the state trajectory, not of individual records\. This paper envisions Governed Evolving Memory, an abstraction that enforces this property through four state\-level operators and six correctness conditions\. Our MemState prototype realizes the abstraction on a property\-graph substrate and exposes what a native engine must deliver\. Three directions define memory\-centric data management as a workload: a native engine for governed memory, trajectory\-level correctness and evaluation, and privacy under shared salience\. The vision succeeds when long\-term memory joins transactions and streams as a recognized data\-management workload\.

## References

- R\. Angles and C\. Gutierrez \(2008\)Survey of graph database models\.ACM Computing Surveys \(CSUR\)40\(1\),pp\. 1–39\.External Links:[Link](https://dl.acm.org/doi/pdf/10.1145/1322432.1322433)Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1)\.
- Anthropic \(2025a\)Claude code\.Anthropic\.External Links:[Link](https://www.anthropic.com/claude-code)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p2.1)\.
- Anthropic \(2025b\)Claude\.Anthropic\.External Links:[Link](https://www.anthropic.com/)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p2.1)\.
- Anysphere \(2025\)Cursor\.Anysphere\.External Links:[Link](https://cursor.com/)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p2.1)\.
- S\. Ceri and J\. Widom \(1990\)Deriving production rules for constraint maintenance\.InProceedings of the International Conference on Very Large Data Bases \(VLDB\),pp\. 566–577\.Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1)\.
- F\. Chang, J\. Dean, S\. Ghemawat, W\. C\. Hsieh, D\. A\. Wallach, M\. Burrows, T\. Chandra, A\. Fikes, and R\. E\. Gruber \(2008\)Bigtable: a distributed storage system for structured data\.ACM Transactions on Computer Systems \(TOCS\)26\(2\),pp\. 1–26\.External Links:[Link](https://dl.acm.org/doi/pdf/10.1145/1365815.1365816)Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1)\.
- P\. Chhikara, D\. Khant, S\. Aryan, T\. Singh, and D\. Yadav \(2025\)Mem0: building production\-ready ai agents with scalable long\-term memory\.arXiv Preprint\.External Links:[Link](https://arxiv.org/pdf/2504.19413)Cited by:[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p3.1),[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p3.1),[§2\.3](https://arxiv.org/html/2605.26252#S2.SS3.p3.1),[§2](https://arxiv.org/html/2605.26252#S2.p3.1),[§5](https://arxiv.org/html/2605.26252#S5.p3.3)\.
- E\. F\. Codd \(1970\)A relational model of data for large shared data banks\.Communications of the ACM13\(6\),pp\. 377–387\.External Links:[Link](https://dl.acm.org/doi/pdf/10.1145/362384.362685)Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1)\.
- CrewAI Inc\. \(2025\)CrewAI: a framework for building role\-based multi\-agent systems with llms\.Note:[https://www\.crewai\.com/](https://www.crewai.com/)Accessed: 2026Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1)\.
- G\. DeCandia, D\. Hastorun, M\. Jampani, G\. Kakulapati, A\. Lakshman, A\. Pilchin, S\. Sivasubramanian, P\. Vosshall, and W\. Vogels \(2007\)Dynamo: amazon’s highly available key\-value store\.ACM SIGOPS Operating Systems Review41\(6\),pp\. 205–220\.External Links:[Link](https://dl.acm.org/doi/pdf/10.1145/1323293.1294281)Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1)\.
- R\. Elmasri and S\. Navathe \(2016\)Fundamentals of database systems\.Pearson\.Cited by:[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p2.1)\.
- R\. Fagin, P\. G\. Kolaitis, R\. J\. Miller, and L\. Popa \(2005\)Data exchange: semantics and query answering\.Theoretical Computer Science\.Cited by:[§5](https://arxiv.org/html/2605.26252#S5.p4.5)\.
- X\. Feng, G\. Jin, Z\. Chen, C\. Liu, and S\. Salihoğlu \(2023\)Kùzu graph database management system\.InConference on Innovative Data Systems Research \(CIDR\),External Links:[Link](https://vldb.org/cidrdb/papers/2023/p48-jin.pdf)Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§4\.1](https://arxiv.org/html/2605.26252#S4.SS1.p1.1)\.
- N\. Francis, A\. Green, P\. Guagliardo, L\. Libkin, T\. Lindaaker, V\. Marsault, S\. Plantikow, M\. Rydberg, P\. Selmer, and A\. Taylor \(2018\)Cypher: an evolving query language for property graphs\.InProceedings of the ACM SIGMOD International Conference on Management of Data \(SIGMOD\),pp\. 1433–1445\.External Links:[Link](https://dl.acm.org/doi/pdf/10.1145/3183713.3190657)Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1),[§5](https://arxiv.org/html/2605.26252#S5.p2.1)\.
- C\. Hu, X\. Gao, Z\. Zhou, D\. Xu, and et al\. \(2026\)EverMemOS: a self\-organizing memory operating system for structured long\-horizon reasoning\.arXiv preprint arXiv:2601\.02163\.External Links:[Link](https://arxiv.org/abs/2601.02163)Cited by:[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p3.1),[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p3.1),[§2\.3](https://arxiv.org/html/2605.26252#S2.SS3.p3.1),[§2](https://arxiv.org/html/2605.26252#S2.p3.1)\.
- Y\. Hu, Y\. Wang, and J\. McAuley \(2025\)Evaluating memory in LLM agents via incremental multi\-turn interactions\.InProceedings of the ICML 2025 Workshop on Long\-Context Foundation Models \(ICML\),External Links:[Link](https://openreview.net/forum?id=ZgQ0t3zYTQ)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1),[§1](https://arxiv.org/html/2605.26252#S1.p5.1),[§1](https://arxiv.org/html/2605.26252#S1.p8.1),[§5](https://arxiv.org/html/2605.26252#S5.p3.3)\.
- C\. S\. Jensen and R\. T\. Snodgrass \(2002\)Temporal data management\.IEEE Transactions on Knowledge and Data Engineering11\(1\),pp\. 36–44\.External Links:[Link](https://vbn.aau.dk/ws/files/310302702/tdb_tutorial_ed_csj_4_uncommented.pdf)Cited by:[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p2.1),[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2\.3](https://arxiv.org/html/2605.26252#S2.SS3.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1)\.
- H\. Kondylakis, S\. Dumbrava, M\. Lissandrini, N\. Yakovets, A\. Bonifati, V\. Efthymiou, G\. Fletcher, D\. Plexousakis, R\. Tommasini, G\. Troullinou,et al\.\(2025\)Property graph standards: state of the art and open challenges\.Proc\. VLDB Endowment \(PVLDB\)\.Cited by:[§5](https://arxiv.org/html/2605.26252#S5.p2.1)\.
- LangChain Inc\. \(2026\)LangGraph: a library for building multi\-agent workflows with llms\.Note:[https://docs\.langchain\.com/oss/python/langgraph/](https://docs.langchain.com/oss/python/langgraph/)Accessed: 2026Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1),[§1](https://arxiv.org/html/2605.26252#S1.p2.1)\.
- Z\. Li, S\. Song, C\. Xi, H\. Wang, C\. Tang, S\. Niu, D\. Chen, Q\. Yang, P\. Yu, and J\. Huo \(2025\)MemOS: a memory os for ai system\.arXiv Preprint\.External Links:[Link](https://arxiv.org/pdf/2507.03724)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p3.1),[§2\.3](https://arxiv.org/html/2605.26252#S2.SS3.p3.1),[§2](https://arxiv.org/html/2605.26252#S2.p3.1)\.
- A\. Maharana, D\. Lee, S\. Tulyakov, and M\. Bansal \(2024\)Evaluating very long\-term conversational memory of LLM agents\.InProceedings of the Annual Meeting of the Association for Computational Linguistics \(ACL\),pp\. 13851–13870\.External Links:[Link](https://aclanthology.org/2024.acl-long.747.pdf)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p6.1),[§5](https://arxiv.org/html/2605.26252#S5.p3.3)\.
- Y\. A\. Malkov and D\. A\. Yashunin \(2018\)Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs\.IEEE Transactions on Pattern Analysis and Machine Intelligence42\(4\),pp\. 824–836\.External Links:[Link](https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8594636)Cited by:[§2\.4](https://arxiv.org/html/2605.26252#S2.SS4.p2.1),[§5](https://arxiv.org/html/2605.26252#S5.p2.1)\.
- T\. Neumann and G\. Weikum \(2010\)X\-rdf\-3x: fast querying, high update rates, and consistency for rdf databases\.Proceedings of the VLDB Endowment \(PVLDB\)3\(1\-2\),pp\. 256–263\.External Links:[Link](https://vldb.org/pvldb/vol3/R22.pdf)Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1)\.
- D\. Olteanu \(2024\)Recent increments in incremental view maintenance\.InPODS,pp\. 12–25\.External Links:[Document](https://dx.doi.org/10.1145/3635138.3654763)Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1)\.
- R\. Omar, I\. Dhall, P\. Kalnis, and E\. Mansour \(2023\)A universal question\-answering platform for knowledge graphs\.Proceedings of the ACM on Management of Data \(SIGMOD\)1\(1\),pp\. 1–25\.External Links:[Link](https://dl.acm.org/doi/pdf/10.1145/3588696)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1)\.
- R\. Omar, A\. Orogat, I\. Abdelaziz, O\. Mangukiya, P\. Kalnis, and E\. Mansour \(2026\)Chatty\-kg: a multi\-agent ai system for on\-demand conversational question answering over knowledge graphs\.Proceedings of the ACM on Management of Data \(SIGMOD\)\.External Links:[Link](https://dl.acm.org/doi/abs/10.1145/3786632)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1)\.
- OpenAI \(2025a\)ChatGPT\.OpenAI\.External Links:[Link](https://chat.openai.com/)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p2.1)\.
- OpenAI \(2025b\)OpenAI agents sdk: a python framework for building and orchestrating multi\-agent systems\.Note:[https://openai\.github\.io/openai\-agents\-python/](https://openai.github.io/openai-agents-python/)Accessed: Nov\. 2025Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1),[§1](https://arxiv.org/html/2605.26252#S1.p2.1)\.
- C\. Packer, S\. Wooders, K\. Lin, V\. Fang, S\. G\. Patil, I\. Stoica, and J\. E\. Gonzalez \(2023\)MemGPT: towards LLMs as operating systems\.arXiv Preprint\.External Links:[Link](https://arxiv.org/pdf/2310.08560)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p3.1),[§2\.3](https://arxiv.org/html/2605.26252#S2.SS3.p3.1),[§2](https://arxiv.org/html/2605.26252#S2.p3.1)\.
- J\. J\. Pan, J\. Wang, and G\. Li \(2024a\)Survey of vector database management systems\.The VLDB Journal33\(5\),pp\. 1591–1615\.External Links:[Link](https://doi.org/10.1007/s00778-024-00864-x)Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1)\.
- J\. J\. Pan, J\. Wang, and G\. Li \(2024b\)Vector database management techniques and systems\.InProceedings of the ACM SIGMOD International Conference on Management of Data \(SIGMOD\),pp\. 597–604\.External Links:[Link](https://dl.acm.org/doi/pdf/10.1145/3626246.3654691)Cited by:[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p2.1)\.
- J\. S\. Park, J\. C\. O’Brien, C\. J\. Cai, M\. R\. Morris, P\. Liang, and M\. S\. Bernstein \(2023\)Generative agents: interactive simulacra of human behavior\.InProceedings of the ACM Symposium on User Interface Software and Technology \(UIST\),External Links:[Link](https://arxiv.org/pdf/2304.03442)Cited by:[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p3.1),[§2\.3](https://arxiv.org/html/2605.26252#S2.SS3.p3.1),[§2\.4](https://arxiv.org/html/2605.26252#S2.SS4.p3.1),[§2](https://arxiv.org/html/2605.26252#S2.p1.1),[§2](https://arxiv.org/html/2605.26252#S2.p3.1)\.
- J\. Perez, M\. Arenas, and C\. Gutierrez \(2009\)Semantics and complexity of sparql\.ACM Transactions on Database Systems \(TODS\)34\(3\),pp\. 1–45\.External Links:[Link](https://dl.acm.org/doi/pdf/10.1145/1567274.1567278)Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1),[§5](https://arxiv.org/html/2605.26252#S5.p2.1)\.
- P\. Rasmussen, P\. Paliychuk, T\. Beauvais, J\. Ryan, and D\. Chalef \(2025\)Zep: a temporal knowledge graph architecture for agent memory\.arXiv Preprint\.External Links:[Link](https://arxiv.org/pdf/2501.13956)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1),[§1](https://arxiv.org/html/2605.26252#S1.p2.1),[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p3.1),[§2\.3](https://arxiv.org/html/2605.26252#S2.SS3.p3.1),[§2\.4](https://arxiv.org/html/2605.26252#S2.SS4.p3.1),[§2](https://arxiv.org/html/2605.26252#S2.p3.1),[§5](https://arxiv.org/html/2605.26252#S5.p3.3)\.
- R\. T\. Snodgrass \(1999\)Developing time\-oriented database applications in sql\.Morgan Kaufmann Publishers\.Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2\.3](https://arxiv.org/html/2605.26252#S2.SS3.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1)\.
- J\. Sun, G\. Li, J\. Pan, J\. Wang, and et al\. \(2025\)GaussDB\-vector: a large\-scale persistent real\-time vector database for llm applications\.Proc\. VLDB Endowment \(PVLDB\)\.External Links:[Link](https://www.vldb.org/pvldb/vol18/p4951-sun.pdf)Cited by:[§5](https://arxiv.org/html/2605.26252#S5.p2.1)\.
- H\. Tan, Z\. Zhang, C\. Ma, X\. Chen, and et al\. \(2025\)MemBench: towards more comprehensive evaluation on the memory of llm\-based agents\.InFindings of the Association for Computational Linguistics \(ACL\),pp\. 19336–19352\.External Links:[Link](https://aclanthology.org/2025.findings-acl.989.pdf)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1),[§1](https://arxiv.org/html/2605.26252#S1.p5.1),[§1](https://arxiv.org/html/2605.26252#S1.p8.1),[§5](https://arxiv.org/html/2605.26252#S5.p3.3)\.
- J\. Wang, X\. Yi, R\. Guo, H\. Jin, P\. Xu, S\. Li, X\. Wang, X\. Guo, C\. Li, X\. Xu,et al\.\(2021\)Milvus: a purpose\-built vector data management system\.InProceedings of the ACM SIGMOD International Conference on Management of Data \(SIGMOD\),pp\. 2614–2627\.External Links:[Link](https://dl.acm.org/doi/pdf/10.1145/3448016.3457550)Cited by:[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p2.1),[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1),[§2\.3](https://arxiv.org/html/2605.26252#S2.SS3.p2.1),[§2](https://arxiv.org/html/2605.26252#S2.p2.1)\.
- Y\. Wang and X\. Chen \(2025\)Mirix: multi\-agent memory system for llm\-based agents\.arXiv Preprint\.External Links:[Link](https://arxiv.org/pdf/2507.07957)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1),[§1](https://arxiv.org/html/2605.26252#S1.p2.1),[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p3.1),[§2](https://arxiv.org/html/2605.26252#S2.p3.1)\.
- Y\. Wang, R\. Takanobu, Z\. Liang, Y\. Mao, and et al\. \(2025\)Mem\-α\\alpha: learning memory construction via reinforcement learning\.arXiv preprint arXiv:2509\.25911\.External Links:[Link](https://arxiv.org/pdf/2509.25911)Cited by:[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p3.1),[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p3.1),[§2](https://arxiv.org/html/2605.26252#S2.p3.1),[§5](https://arxiv.org/html/2605.26252#S5.p3.3)\.
- J\. Widom and S\. Ceri \(1996\)Active database systems: triggers and rules for advanced database processing\.Cited by:[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p2.1)\.
- D\. Wu, H\. Wang, W\. Yu, Y\. Zhang, K\. Chang, and D\. Yu \(2025\)LongMemEval: benchmarking chat assistants on long\-term interactive memory\.InProceedings of the International Conference on Learning Representations \(ICLR\),External Links:[Link](https://openreview.net/forum?id=pZiyCaVuti)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1),[§1](https://arxiv.org/html/2605.26252#S1.p7.1),[§5](https://arxiv.org/html/2605.26252#S5.p3.3)\.
- W\. Xu, Z\. Liang, K\. Mei, H\. Gao, J\. Tan, and Y\. Zhang \(2025\)A\-mem: agentic memory for LLM agents\.InProceedings of the Conference on Neural Information Processing Systems \(NeurIPS\),External Links:[Link](https://openreview.net/pdf?id=FiM0M8gcct)Cited by:[§1](https://arxiv.org/html/2605.26252#S1.p1.1),[§1](https://arxiv.org/html/2605.26252#S1.p2.1)\.
- S\. Yan, X\. Yang, Z\. Huang, E\. Nie, and et al\. \(2025\)Memory\-r1: enhancing large language model agents to manage and utilize memories via reinforcement learning\.CoRR abs/2508\.19828\.External Links:[Link](https://doi.org/10.48550/arXiv.2508.19828)Cited by:[§2\.1](https://arxiv.org/html/2605.26252#S2.SS1.p3.1),[§2\.2](https://arxiv.org/html/2605.26252#S2.SS2.p3.1),[§2](https://arxiv.org/html/2605.26252#S2.p3.1),[§5](https://arxiv.org/html/2605.26252#S5.p3.3)\.