Beyond Retrieval: Analytic Memory for Multimodal Agents

# Beyond Retrieval: Analytic Memory for Multimodal Agents
Source: [https://arxiv.org/html/2607.29440](https://arxiv.org/html/2607.29440)
Zhoujin Tian1,Yao Tian2,Hao Zhang2,Cheng Chen2 Yakun Li2,Lei Zhang2,Xiaofang Zhou1 1HKUST,2ByteDance \{ztianaf, zxf\}@cse\.ust\.hk \{yao\.tian, zhanghao\.ai, chencheng\.sg, liyakun\.hit, zhanglei\.michael\}@bytedance\.com

###### Abstract

Long\-term multimodal memory must support not only retrieving relevant information but also computing over observations accumulated across interactions\. Existing systems largely emphasize*retrieval memory*, organizing interaction histories through summaries and indexes to return query\-relevant information at multiple granularities, from high\-level abstractions to underlying records\. In this paper, we formulate*analytic memory*as a complementary abstraction that organizes recurring multimodal observations into queryable structures supporting filtering, aggregation, ranking, and temporal comparison\. We presentAdaMM, a framework that jointly supports retrieval and analytic memory\. Rather than relying on application\-defined schemas,AdaMMextracts provenance\-linked attribute\-value observations from dialogue, images, and contextual metadata, discovers recurring field structures, and materializes them for analytical access\. At inference time, a memory\-aware planner decomposes queries into retrieval and analytic operations and routes each operation to the appropriate tools\. Experiments on two long\-term multimodal memory benchmarks, MemEye and MemGallery, show thatAdaMMimproves performance by up to 11\.3% and 7\.3%, respectively\.

Beyond Retrieval: Analytic Memory for Multimodal Agents

Zhoujin Tian1, Yao Tian2††thanks:Corresponding author\., Hao Zhang2, Cheng Chen2Yakun Li2,Lei Zhang2,Xiaofang Zhou11HKUST,2ByteDance\{ztianaf, zxf\}@cse\.ust\.hk\{yao\.tian, zhanghao\.ai, chencheng\.sg, liyakun\.hit, zhanglei\.michael\}@bytedance\.com

## 1Introduction

As Large Language Model \(LLM\) agents interact with users over increasingly long horizons, they accumulate extensive multimodal histories – including dialogues, screenshots, images, documents, and contextual observations – that quickly exceed the finite context windows of their backbone models\(Heet al\.,[2024](https://arxiv.org/html/2607.29440#bib.bib38); Kohet al\.,[2024](https://arxiv.org/html/2607.29440#bib.bib39); Xieet al\.,[2024](https://arxiv.org/html/2607.29440#bib.bib40); Wanget al\.,[2024](https://arxiv.org/html/2607.29440#bib.bib41)\)\. Multimodal memory systems, which retain and organize such histories to support future reasoning\(Wang and Chen,[2025](https://arxiv.org/html/2607.29440#bib.bib33); Fenget al\.,[2026](https://arxiv.org/html/2607.29440#bib.bib34)\), have therefore become essential and are widely adopted in applications such as personal assistance\(Liet al\.,[2024](https://arxiv.org/html/2607.29440#bib.bib22); Fenget al\.,[2026](https://arxiv.org/html/2607.29440#bib.bib34)\), embodied interaction\(Driesset al\.,[2023](https://arxiv.org/html/2607.29440#bib.bib42); Wanget al\.,[2023](https://arxiv.org/html/2607.29440#bib.bib43)\), and tool\-augmented workflows\(Yaoet al\.,[2022](https://arxiv.org/html/2607.29440#bib.bib44); Schicket al\.,[2023](https://arxiv.org/html/2607.29440#bib.bib45)\)\.

![Refer to caption](https://arxiv.org/html/2607.29440v1/x1.png)Figure 1:Long\-term multimodal interaction histories contain observations distributed across sessions, modalities, and time\. Answering user queries may require either retrieving specific past events or performing analytical operations, such as temporal selection and aggregation, over multiple observations\.Recent multimodal memory systems have explored a broad range of designs, including textual abstraction\(Fenget al\.,[2026](https://arxiv.org/html/2607.29440#bib.bib34)\), specialized memory types\(Nguyenet al\.,[2026](https://arxiv.org/html/2607.29440#bib.bib36)\), hybrid stores\(Wang and Chen,[2025](https://arxiv.org/html/2607.29440#bib.bib33)\), and cross\-modal retrieval\(Luet al\.,[2026](https://arxiv.org/html/2607.29440#bib.bib35)\)\. Most existing systems follows a retrieve\-then\-answer paradigm, conditioning an LLM on a bounded set of relevant memories, which we denote as retrieval memory\. Such systems are effective at selecting relevant memories from long interaction histories\. However, long interaction histories also accumulate recurring observations that collectively form an append\-only log\. Analytical questions over such histories require complete, correctly scoped records and operations such as filtering, aggregation, ranking, and temporal selection\. Relevance\-based approach adopted by retrieval memory systems therefore creates a coverage–context trade\-off: a small retrieval set may omit required observations, whereas a larger retrieval set consumes the limited context budget and introduces redundant or distracting evidence\. For example, as illustrated in Figure[1](https://arxiv.org/html/2607.29440#S1.F1), computing a user’s average sleep duration over the past month requires collecting all relevant measurements across multiple screenshots before aggregation; incomplete retrieval may therefore produce a biased estimate\. We call this gap between relevance\-based retrieval and analytic the*retrieval–analysis mismatch*\.

To bridge this gap, our key insight is that a memory system should couple two complementary subsystems:*retrieval memory*, which identifies and supplies relevant historical records as context, and*analytic memory*, which proactively organizes interaction histories into reusable and queryable structures and adaptively executes operations over them\. Together, they support both flexible semantic recall and structured analysis, rather than requiring the LLM to repeatedly reconstruct the needed organization from retrieved records\. Realizing analytic memory, however, presents two key challenges\. First, interaction histories do not come with an explicit organizing schema, making it unclear what information should be preserved and how they should be structured to support future queries\. Second, user queries rarely specify memory access strategies, the system therefore must reconcile what each query requires with what the evolving memory can currently support\. Even questions over the same underlying records may require fundamentally different operations \(e\.g\., aggregation, ordering, filtering\), posing challenges to the design of adaptive execution plans\.

To address these challenges, we introduceAdaMM, anAdaptiveMulti\-viewMemory framework that couples data\-driven structure induction with memory\-aware query planning\. It maintains two complementary subsystems:*retrieval memory*for flexible semantic access and*analytic memory*for executable analysis over recurring observations\.AdaMMextracts provenance\-linked attribute–value pairs from multimodal interactions \(e\.g\., \{Sleep Time: 5\.5 hrs\}\), induces recurrent schemas, and materializes them as structured tables, which is accessible via tools, while preserving unstructured evidence in a hierarchical semantic graph\. At query time, a planner adaptively composes different analytic and retrieval tools based on the query and the memory structures currently available\. This design enables grounded semantic recall and structured analysis over long\-term interaction histories\.

Our contributions are summarized as follows:

- •We identify a*retrieval–analysis mismatch*in long\-term multimodal memory: retrieval\-oriented interfaces support returning query\-relevant information at multiple granularities, but cannot reliably answer queries requiring complete\-range filtering, aggregation, ranking, or temporal selection over append\-only interaction histories\.
- •We proposeAdaMM, a multimodal memory framework that complements retrieval memory with schema\-induced analytic memory\.AdaMMdiscovers recurrent structures from multimodal observations, materializes them as queryable tables, and uses a memory\-aware planner to select or compose semantic retrieval with native analytical operations\.
- •We evaluateAdaMMon two multimodal long\-term memory benchmarks, MemEye and MemGallery, whereAdaMMconsistently outperforms strong memory baselines, and improves accuracy by up to 11\.3% and 7\.3%, respectively\.

## 2Related Work

##### Agent memory\.

Long\-term agent memory preserves information across interactions, allowing agents to reuse preferences, observations, decisions, and task states beyond a finite context windowHataliset al\.\([2023](https://arxiv.org/html/2607.29440#bib.bib21)\); Liet al\.\([2024](https://arxiv.org/html/2607.29440#bib.bib22)\); Zhanget al\.\([2025](https://arxiv.org/html/2607.29440#bib.bib28)\)\. Existing systems explore personalized stores and tiered memory managementZhonget al\.\([2024](https://arxiv.org/html/2607.29440#bib.bib6)\); Packeret al\.\([2023](https://arxiv.org/html/2607.29440#bib.bib37)\), scalable or agentic memory constructionChhikaraet al\.\([2025](https://arxiv.org/html/2607.29440#bib.bib15)\); Xuet al\.\([2025](https://arxiv.org/html/2607.29440#bib.bib17)\), and temporal or relational organization[Rasmussenet al\.](https://arxiv.org/html/2607.29440#bib.bib16); Huanget al\.\([2025](https://arxiv.org/html/2607.29440#bib.bib46)\)\. Recent schema\-grounded memory moves beyond semantic recall by transforming textual interactions into validated, queryable records, but relies on application\-defined schemasPetrovet al\.\([2026](https://arxiv.org/html/2607.29440#bib.bib3)\)\. Despite these advances, existing methods remain largely text\-centric, leaving underexplored how to induce reusable analytical structures from heterogeneous multimodal histories and adaptively execute retrieval and analytical operations over them\.

##### Multimodal agent memory\.

Multimodal agent memory extends text\-centric memory to histories containing heterogeneous visual observations\. Existing approaches broadly follow two directions: MIRIX and M2A translate visual content into captions, summaries, or semantic abstractions for text\-based retrievalWang and Chen \([2025](https://arxiv.org/html/2607.29440#bib.bib33)\); Fenget al\.\([2026](https://arxiv.org/html/2607.29440#bib.bib34)\), whereas MMA, VisualMem, and Omni\-SimpleMem retain native visual evidence and retrieve multimodal or image\-backed memoriesLuet al\.\([2026](https://arxiv.org/html/2607.29440#bib.bib35)\); Nguyenet al\.\([2026](https://arxiv.org/html/2607.29440#bib.bib36)\); Liuet al\.\([2026](https://arxiv.org/html/2607.29440#bib.bib50)\)\. Although these methods improve visual evidence preservation, they still primarily organize histories as semantic memories\. MemEye further reveals their difficulty in capturing fine\-grained details and evolving visual statesGuoet al\.\([2026](https://arxiv.org/html/2607.29440#bib.bib47)\)\. In contrast,AdaMMadaptively organizes multimodal histories into complementary analytical and semantic memories, enabling both precise analysis and context\-aware retrieval for complex queries\.

## 3Method

![Refer to caption](https://arxiv.org/html/2607.29440v1/x2.png)Figure 2:Overview ofAdaMM\. It complements recall\-oriented retrieval memory with schema\-induced analytic memory over recurrent multimodal observations\. At query time, an operation planner jointly considers the query and the instantiated memories to select designed access tools for answer generation\.### 3\.1Problem Formulation

Letℋ=\{Si\}i=1M\\mathcal\{H\}=\\\{S\_\{i\}\\\}\_\{i=1\}^\{M\}denote a multimodal interaction history, where each sessionSi=\{Ri,j\}j=1niS\_\{i\}=\\\{R\_\{i,j\}\\\}\_\{j=1\}^\{n\_\{i\}\}consists of multiple interaction rounds\. Each roundRi,j=\(di,j,𝒱i,j,τi,j\)R\_\{i,j\}=\(d\_\{i,j\},\\mathcal\{V\}\_\{i,j\},\\tau\_\{i,j\}\)comprises the user–agent dialoguedi,jd\_\{i,j\}, visual observations𝒱i,j\\mathcal\{V\}\_\{i,j\}, and temporal informationτi,j\\tau\_\{i,j\}\. A multimodal agent memory system transforms the interaction history into memory representations,ℳ=Fbuild​\(ℋ\),\\mathcal\{M\}=F\_\{\\mathrm\{build\}\}\(\\mathcal\{H\}\),and aims to generate an accurate answery^=Fanswer​\(ℳ,q\)\\hat\{y\}=F\_\{\\mathrm\{answer\}\}\(\\mathcal\{M\},q\)for a user queryqqbased onℳ\\mathcal\{M\}\.

### 3\.2Overview

Figure[2](https://arxiv.org/html/2607.29440#S3.F2)presents the overview of the proposedAdaMMframework\. Given a long\-term multimodal interaction history,AdaMMconstructs two complementary memories: analytic memoryℳana\\mathcal\{M\}^\{\\mathrm\{ana\}\}induces recurring patterns as adaptive schemas for analytical queries while retrieval memoryℳret\\mathcal\{M\}^\{\\mathrm\{ret\}\}hierarchically organizes interactions according to their semantics to enable flexible retrieval\. At query time,AdaMMjointly reasons over the query and the memories instantiated from the current history, then constructs an execution plan by adaptively selecting and composing access tools to gather sufficient evidence for grounded answer generation\.

### 3\.3Analytic Memory Construction

Analytic memory aims to uncover and organize recurring patterns across fragmented observations in multimodal interaction histories for analytical operations\. The key challenge is that the meaningful attributes, which is in the form of key\-value pair, and their co\-occurrence patterns are unknown in advance, and each round reveals only a sparse fragment of the latent structure\. To address this challenge,AdaMMtreatsℋ\\mathcal\{H\}as a chronologically ordered interaction rounds\{Rt\}t=1N\\\{R\_\{t\}\\\}\_\{t=1\}^\{N\}, performs*Attribute Extraction*to adaptively identify attribute evidence within each round, then conducts*Schema Induction*to discover stable structures from recurring cross\-round patterns, and finally applies*Memory Materialization*to transform the induced schemas into executable analytic memory\.

#### 3\.3\.1Attributes Extraction

Recovering recurring structure must begin from the evidence available within individual rounds, yet the attributes that will form a useful schema are unknown at extraction time\. Record fragment extraction therefore aims to preserve explicit analytic observations from each round without imposing a predefined and unobserved pattern\.

GivenRtR\_\{t\},AdaMMjointly examines its dialogue and visual content to identify every grounded attribute–value correspondence:

𝒪t=Extractor⁡\(Rt\)=\{\(at​ℓ,xt​ℓ,pt​ℓ\)\}ℓ=1mt,\\mathcal\{O\}\_\{t\}=\\operatorname\{Extractor\}\(R\_\{t\}\)=\\left\\\{\\left\(a\_\{t\\ell\},x\_\{t\\ell\},p\_\{t\\ell\}\\right\)\\right\\\}\_\{\\ell=1\}^\{m\_\{t\}\},\(1\)whereat​ℓa\_\{t\\ell\}andxt​ℓx\_\{t\\ell\}are the observed attribute and its source\-faithful value, andpt​ℓp\_\{t\\ell\}points to the supporting dialogue span or image\.Extractor\\operatorname\{Extractor\}is an LLM\-based extractor that targets at identifying all possible pairs from each round\. The resulting collection\{𝒪t\}t=1N\\\{\\mathcal\{O\}\_\{t\}\\\}\_\{t=1\}^\{N\}captures these observations and their within\-interaction co\-occurrence patterns for subsequent schema induction\.

#### 3\.3\.2Schema Induction

Given the extracted record fragments, it remains challenging to induce reusable schemas without assuming a predefined schema\. New observations may reveal either a previously unseen structure or evolution to an existing one\. To address this challenge,AdaMMfirst mine candidate patterns and then induces schemas through two processes:*Schema Discovery*creates new schemas, while*Schema Evolution*extends existing ones with consistently co\-occurring attributes\.

##### Candidate Pattern Mining\.

For each extracted record fragment𝒪t\\mathcal\{O\}\_\{t\}, let

𝒜t=\{at​ℓ:\(at​ℓ,xt​ℓ,pt​ℓ\)∈𝒪t\}\\mathcal\{A\}\_\{t\}=\\left\\\{a\_\{t\\ell\}:\\left\(a\_\{t\\ell\},x\_\{t\\ell\},p\_\{t\\ell\}\\right\)\\in\\mathcal\{O\}\_\{t\}\\right\\\}\(2\)denote the set of attributes contained in𝒪t\\mathcal\{O\}\_\{t\}\. A reliable candidate patternCCis a set of attributes that consistently recur sufficient interactions together\. Following the Apriori framework for frequent\-itemset mining\(Agrawal and Srikant,[1994](https://arxiv.org/html/2607.29440#bib.bib4)\), we use*support*to quantify information recurrence\. Specifically, the*support*of a candidate pattern𝒞\\mathcal\{C\}after processing roundttis

suppt⁡\(𝒞\)=∑s=1t𝕀​\[C⊆𝒜s\],\\operatorname\{supp\}\_\{t\}\(\\mathcal\{C\}\)=\\sum\_\{s=1\}^\{t\}\\mathbb\{I\}\\\!\\left\[C\\subseteq\\mathcal\{A\}\_\{s\}\\right\],\(3\)AdaMMretains𝒞\\mathcal\{C\}as a candidate pattern if

suppt⁡\(𝒞\)≥θs,\\operatorname\{supp\}\_\{t\}\(\\mathcal\{C\}\)\\geq\\theta\_\{s\},\(4\)which ensures that𝒞\\mathcal\{C\}recurs across sufficient interaction rounds to yield a populated schema\.

##### Schema Discovery\.

For each recurrent candidate𝒞\\mathcal\{C\},AdaMMfirst compares it with the active schemas before roundtt\. If𝒞\\mathcal\{C\}exhibits an attribute structure distinct from all existing schemas, it is evaluated for reliable new\-schema discovery\. Following a variant of Apriori algorithm\(Omiecinski,[2003](https://arxiv.org/html/2607.29440#bib.bib1)\), we use*all\-confidence*to measure within\-pattern co\-occurrence consistency, defined as

all−conft⁡\(𝒞\)=suppt⁡\(𝒞\)maxa∈𝒞⁡suppt⁡\(\{a\}\)\.\\operatorname\{all\-conf\}\_\{t\}\(\\mathcal\{C\}\)=\\frac\{\\operatorname\{supp\}\_\{t\}\(\\mathcal\{C\}\)\}\{\\max\_\{a\\in\\mathcal\{C\}\}\\operatorname\{supp\}\_\{t\}\(\\\{a\\\}\)\}\.\(5\)A candidate is admitted as a new schema if

all−conft⁡\(𝒞\)≥θa\.\\operatorname\{all\-conf\}\_\{t\}\(\\mathcal\{C\}\)\\geq\\theta\_\{a\}\.\(6\)This constraint requires the complete candidate to occur reliably whenever any of its constituent attributes occurs, thereby preventing a ubiquitous attributes from being merged with most patterns\. To avoid admitting a qualified candidate together with its qualified subsets,AdaMMretains only inclusion\-maximal novel candidates\.

##### Schema Evolution\.

For an existing schema𝒜\\mathcal\{A\}, if𝒞⊆𝒜\\mathcal\{C\}\\subseteq\\mathcal\{A\}, then𝒞\\mathcal\{C\}is already subsumed by it\. If𝒞\\mathcal\{C\}contains𝒜\\mathcal\{A\},AdaMMtreats𝒞\\mathcal\{C\}as a potential extension of𝒜\\mathcal\{A\}\. LetΔ=𝒞∖𝒜\\Delta=\\mathcal\{C\}\\setminus\\mathcal\{A\}is the newly added attributes\. We use*extension confidence*to measure how consistentlyΔ\\Deltaaccompany occurrences of𝒜\\mathcal\{A\}, defined as

ext−conft⁡\(𝒞∣𝒜\)=suppt⁡\(𝒞\)suppt⁡\(𝒜\)\\operatorname\{ext\-conf\}\_\{t\}\(\\mathcal\{C\}\\mid\\mathcal\{A\}\)=\\frac\{\\operatorname\{supp\}\_\{t\}\(\\mathcal\{C\}\)\}\{\\operatorname\{supp\}\_\{t\}\(\\mathcal\{A\}\)\}\(7\)Whenext−conft⁡\(𝒞∣𝒜\)≥θe\\operatorname\{ext\-conf\}\_\{t\}\(\\mathcal\{C\}\\mid\\mathcal\{A\}\)\\geq\\theta\_\{e\}, the candidate provides sufficient evidence thatΔ\\Deltahave become stable components of𝒜\\mathcal\{A\}, and we update the schema𝒜\\mathcal\{A\}to𝒞\\mathcal\{C\}\. Otherwise,𝒜\\mathcal\{A\}remains unchanged\.

#### 3\.3\.3Memory Materialization

Considering that each schema specifies a stable set of attributes and each matched interaction provides their co\-occurring values, the induced patterns naturally align with relational tables\.AdaMMtherefore materialize each schema𝒜\\mathcal\{A\}into a tableTAT\_\{A\}, using its attributes as columns and the matched interactions as rows\.

Specifically, for a newly discovered schema𝒜\\mathcal\{A\}, its attributes first define the data columns\. Given these attributes and their sampled values from the corresponding extracted observations\{𝒪t\}\\\{\\mathcal\{O\}\_\{t\}\\\}, a language model generates the table metadata, including its name and description, and infers the type of each column\. We then populate the table at the interaction level\. For each associated roundRtR\_\{t\}, every extracted tuple\(at​ℓ,xt​ℓ,pt​ℓ\)∈𝒪t\(a\_\{t\\ell\},x\_\{t\\ell\},p\_\{t\\ell\}\)\\in\\mathcal\{O\}\_\{t\}whose attribute belongs to𝒜\\mathcal\{A\}is mapped to the corresponding column, and values extracted from the same round jointly form one row\. Two auxiliary columns,orderandtime, record the position and occurrence time of the source interaction, respectively\. Any value not observed in that round is left empty\. When a schema evolves from𝒜\\mathcal\{A\}to𝒜′=𝒜∪Δ\\mathcal\{A\}^\{\\prime\}=\\mathcal\{A\}\\cup\\Delta, the attributes inΔ\\Deltaare appended as new columns to the existing tableT𝒜T\_\{\\mathcal\{A\}\}\.AdaMMthen updates the table information and rematerializes the rows from their associated observations following the same procedure\. The resulting analytic memory is represented as the collection of tablesℳana=\{T\}\\mathcal\{M\}^\{\\mathrm\{ana\}\}=\\\{T\\\}\.

### 3\.4Retrieval Memory Construction

Beyond structured analytic queries, many user requests can be addressed through direct semantic matching against past interactions\. Retrieval memory is designed for such queries by preserving open\-ended events, relations, and visual details that may not exhibit recurring schemas\. It complements analytic memory by enabling flexible semantic retrieval and context\-aware reasoning over interaction histories\.

Following prior work on hierarchical memory organization\(Yueet al\.,[2026](https://arxiv.org/html/2607.29440#bib.bib2)\), we construct retrieval memoryℳr​e​t\\mathcal\{M\}^\{ret\}into three levels:topic→episode→event\\text\{topic\}\\rightarrow\\text\{episode\}\\rightarrow\\text\{event\}\. Events preserve fine\-grained evidence extracted from individual interactions, episodes group temporally adjacent and semantically coherent events, and topics aggregate related episodes into broader semantic contexts\. Each level maintains a concise description, multimodal representation, temporal information, and provenance\. This hierarchy allows the planner to first identify relevant topics or episodes and then locate supporting events, preserving both high\-level context and fine\-grained evidence for complex queries\.

### 3\.5Adaptive Query Processing

Different queries place different demands on memory: some require semantic retrieval over open\-ended events, whereas others require structured computation over analytic memory\. No single retrieval primitive can adequately support this full range of needs, calling for an adaptive mechanism that selects and composes operations according to query intent and the structures available in memory\.AdaMMtherefore abstracts the native capabilities of analytic and semantic memory as operation\-specific tools, and employs a joint query–memory planner to generate an executable procedure for answering each query\.

#### 3\.5\.1Memory Access Tools

Analytic and semantic memory differ in their underlying structures and consequently provide distinct native capabilities\. To expose these capabilities at query time,AdaMMequips each memory with structure\-specific tools\.

Specifically, each tool is represented as

ξi=\(di,Θi,Ωi\),\\xi\_\{i\}=\(d\_\{i\},\\Theta\_\{i\},\\Omega\_\{i\}\),\(8\)wheredid\_\{i\}textually specify its tool function, whileΘi\\Theta\_\{i\}andΩk\\Omega\_\{k\}define its typed arguments and return schema\. Table[1](https://arxiv.org/html/2607.29440#S3.T1)summarizes the access tools designed for analytic and semantic memory\. For example, for the analytic toolCompute,did\_\{i\}specifies a deterministic computation over an induced table,Θi\\Theta\_\{i\}includes the target table, computation operator, target column, and optional constraints, andΩi\\Omega\_\{i\}returns the computed result and matched records\. For the semantic toolSemanticMatch,dkd\_\{k\}specifies relevance\-based memory retrieval,Θi\\Theta\_\{i\}includes a semantic query content and retrieval budget, andΩi\\Omega\_\{i\}returns the ranked memory units with their relevance scores\.

Table 1:Access tools for analytic and retrieval memory\.MemoryToolsCapabilityAnalyticLookup,Filter,Compute,RankExact record access and deterministic computation\.retrievalSemanticMatch,EventLocateRelevance\-based retrieval and event localization\.
#### 3\.5\.2Memory\-Aware Joint Query Planning

The access tools specify how each memory can be queried\. However, users describe desired outcomes rather than access procedures, and a feasible operation also depends on what the current memory contains\. To bridge this gap,AdaMMfirst constructs a query\-conditioned planning context that exposes relevant memory structure, and then employs an LLM\-based planner to jointly reason over the query and this context, composing an executable procedure for retrieving and analyzing the required information\. The resulting evidence is subsequently provided to the answer model for response generation\.

##### Planning Context Construction\.

Valid query planning requires knowledge of the current memory structures and their supported operations\. Exposing the complete memory contents, however, introduces irrelevant information and expands the planning space\.AdaMMtherefore abstracts analytic and retrieval memory into compact metadata, retrieves query\-relevant entries, and associates them with their available tools to form the effective planning context\.

Specifically, we treat each analytic tableT∈ℳanaT\\in\\mathcal\{M\}^\{\\mathrm\{ana\}\}and retrieval episodeP∈ℳretP\\in\\mathcal\{M\}^\{\\mathrm\{ret\}\}as an individual planning candidate represented by textual metadataμT\\mu\_\{T\}andμP\\mu\_\{P\}, respectively\. For each table,μT\\mu\_\{T\}summarizes its name, description, columns, and sampled values, while each episode directly uses its description asμP\\mu\_\{P\}\. Together, these candidates form the metadata search space𝒰\\mathcal\{U\}\. Given a queryqq,AdaMMranks eachμ∈𝒰\\mu\\in\\mathcal\{U\}using

hλ​\(q,μ\)\\displaystyle h\_\{\\lambda\}\(q,\\mu\)=λ​cos⁡\(E​\(q\),E​\(μ\)\)\\displaystyle=\\lambda\\cos\\\!\\left\(E\(q\),E\(\\mu\)\\right\)\(9\)\+\(1−λ\)​TokSim⁡\(q,μ\),\\displaystyle\\quad\+\(1\-\\lambda\)\\operatorname\{TokSim\}\(q,\\mu\),whereE​\(⋅\)E\(\\cdot\)is a semantic encoder andTokSim\\operatorname\{TokSim\}measures token\-level similarity\. After retaining the top\-rankedμ\\mu,AdaMMaugments their metadata with designed access tool descriptions, forming the planning context𝒟q\\mathcal\{D\}\_\{q\}to expose relevant memory candidates and available operations to the planner\.

##### Progressive Tool\-Execution Planning\.

Complex queries often require multiple dependent operations, with the arguments of later tool calls determined by earlier outputs\. Instantiating all tool calls upfront is therefore unreliable\. To address this challenge,AdaMMseparates high\-level planning from progressive tool instantiation\.

*High\-level plan generation\.*Conditioned on the queryqqand planning context𝒟q\\mathcal\{D\}\_\{q\},AdaMMemploys an LLM\-based planner to generate an abstract plan specifying the information goals and their execution order, without binding them to concrete tools or arguments:

π¯​\(q,𝒟q\)=Planner⁡\(q,𝒟q\)=\[\(gℓ,ξℓ\)\]ℓ=1L,\\bar\{\\pi\}\(q,\\mathcal\{D\}\_\{q\}\)=\\operatorname\{Planner\}\(q,\\mathcal\{D\}\_\{q\}\)=\\big\[\(g\_\{\\ell\},\\xi\_\{\\ell\}\)\\big\]\_\{\\ell=1\}^\{L\},\(10\)wheregℓg\_\{\\ell\}is a textual description of the information required at stepℓ\\ell, andξℓ\\xi\_\{\\ell\}specifies the access tool designated to obtain it\.

*Progressive tool instantiation\.*Guided by the above high\-level plan,AdaMMinstantiates tool calls sequentially\. At stepℓ\\ell, it conditions on the preceding results𝒵<ℓ=\{z1,…,zℓ−1\}\\mathcal\{Z\}\_\{<\\ell\}=\\\{z\_\{1\},\\ldots,z\_\{\\ell\-1\}\\\}:

Θℓ\\displaystyle\\Theta\_\{\\ell\}=Planner⁡\(q,gℓ,ξℓ,𝒟q,𝒵<ℓ\),\\displaystyle=\\operatorname\{Planner\}\(q,g\_\{\\ell\},\\xi\_\{\\ell\},\\mathcal\{D\}\_\{q\},\\mathcal\{Z\}\_\{<\\ell\}\),\(11\)zℓ\\displaystyle z\_\{\\ell\}=Invoke⁡\(ξℓ,Θℓ\),\\displaystyle=\\operatorname\{Invoke\}\(\\xi\_\{\\ell\},\\Theta\_\{\\ell\}\),whereξℓ\\xi\_\{\\ell\}andΘℓ\\Theta\_\{\\ell\}denote the selected tool and its instantiated arguments\. This allows later calls to consume intermediate results \(e\.g\., a timestamp returned byEventLocatecan serve as a temporal constraint for a subsequentComputecall\)\.

##### Question Answering\.

Finally, the answer model generates response conditioned on the query and an evidence context comprising the instantiated plan and its tool outputs\.

## 4Experimental Evaluation

### 4\.1Experimental Setup

##### Benchmarks\.

We evaluateAdaMMon two complex multimodal memory benchmarks\.MemEyeevaluates visual\-memory granularity and reasoning through paired multiple\-choice and open\-ended queries\(Guoet al\.,[2026](https://arxiv.org/html/2607.29440#bib.bib47)\), whileMemGalleryassesses long\-term conversational memory management\(Beiet al\.,[2026](https://arxiv.org/html/2607.29440#bib.bib48)\)\.MemEyeuses EM, BLEU\-1 and LLM\-judge for evaluation, whereasMemGalleryreports F1, BLEU\-1, and LLM\-Judge scores\. Dataset statistics are provided in Table[2](https://arxiv.org/html/2607.29440#S4.T2)\.

BenchmarksSessionsRoundsImagesQA pairsMemEye221848438742MemGallery2403,9621,0031,711Table 2:Benchmark statistics\.
##### Baselines\.

We compare against unimodal memory agents, includingA\-MemandMemoryOS\(Xuet al\.,[2025](https://arxiv.org/html/2607.29440#bib.bib17); Liet al\.,[2025](https://arxiv.org/html/2607.29440#bib.bib7)\), which organize long\-term textual interaction histories\. Multimodal baselines include dedicated memory agents \(M2A,MMA, andMIRIX\)\(Fenget al\.,[2026](https://arxiv.org/html/2607.29440#bib.bib34); Luet al\.,[2026](https://arxiv.org/html/2607.29440#bib.bib35); Wang and Chen,[2025](https://arxiv.org/html/2607.29440#bib.bib33)\), which construct persistent memories from dialogue and visual observations, and retrieval\-based systems \(MM\-RAGandUniversalRAG\), which directly retrieve relevant multimodal evidence for answer generation\.

##### Implementation details\.

We use GPT\-4\.1\-nano and GPT\-5\.4\-min as answer and memory construction backbones for all methods\. Text and image representations are produced by MiniLM\-L6\-v2 and siglip2\-base\-patch16\-384, respectively\. We use Top\-10 retrieved memory units for retrieval memory baselines, and our method constrain the planner to generate at most three execution steps under a shared budget of 10 evidence units\. Results are averaged over 3 runs\.

### 4\.2Main Results

BackboneMethodMemEyeMemGalleryEMBLEU\-1LLM\-JudgeF1BLEU\-1LLM\-JudgeGPT\-4\.1\-nanoUnimodalA\-Mem39\.117\.132\.053\.247\.267\.4MemoryOS42\.415\.827\.653\.447\.165\.9MultimodalM2A34\.15\.813\.850\.244\.361\.5MMA39\.010\.632\.055\.950\.764\.1MIRIX40\.95\.713\.656\.351\.166\.3MM\-RAG42\.716\.642\.258\.352\.467\.3UniversalRAG43\.116\.540\.257\.752\.467\.0AdaMM\(Ours\)50\.4\(\+7\.3\)21\.2\(\+4\.1\)48\.0\(\+5\.8\)62\.6\(\+4\.2\)56\.7\(\+4\.3\)74\.3\(\+7\.0\)GPT\-5\.4\-miniUnimodalA\-Mem48\.022\.435\.263\.058\.674\.4MemoryOS48\.721\.833\.364\.559\.075\.5MultimodalM2A40\.112\.333\.560\.856\.071\.7MMA53\.925\.743\.364\.059\.673\.3MIRIX46\.718\.033\.265\.760\.978\.2MM\-RAG61\.829\.749\.466\.561\.978\.7UniversalRAG62\.430\.448\.665\.660\.877\.4AdaMM\(Ours\)65\.5\(\+3\.2\)35\.9\(\+5\.5\)60\.7\(\+11\.3\)69\.1\(\+2\.6\)64\.1\(\+2\.2\)83\.9\(\+5\.2\)Table 3:Main results on MemEye and MemGallery\. Higher is better for all metrics\.bold valuesdenote the best result within a backbone andred valuesdenote improvements over the strongest baseline\.##### Overall performance\.

Table[3](https://arxiv.org/html/2607.29440#S4.T3)presents a comprehensive comparison ofAdaMMwith representative unimodal and multimodal memory frameworks across two benchmarks and answer backbones\.AdaMMconsistently achieves the best result on every metric\. With GPT\-4\.1\-nano, it surpasses the strongest competing result on MemEye by 7\.3% and 5\.8% percentage points for MCQ and open\-ended questions, respectively, and improves MemGallery F1, BLEU\-1, and LLM\-Judge by 4\.2%, 4\.3%, and 7\.0%\. The advantage persists with GPT\-5\.4\-mini, yielding gains of 3\.2% and 11\.3% points on MemEye and 2\.6%, 2\.2%, and 5\.2% on the three MemGallery metrics\. The consistant improvements on different settings demonstrate thatAdaMMgeneralizes across answer formats, benchmark settings, and backbone capacities\.

##### Fine\-grained task analysis\.

To further examine where the improvement arises, Figure[3](https://arxiv.org/html/2607.29440#S4.F3)reports the LLM\-Judge breakdown using GPT\-5\.4\-mini\. On MemEye, the largest gains over the strongest baseline occur on Card Playlog and Personal Health, with improvements of 18\.8% and 16\.7%, respectively\. Both tasks require exact operations over recurring records, such as filtering observations, comparing values, or tracking changes, highlighting the benefit of analytic memory\.AdaMMalso achieves substantial gains on Multi\-scene \(11\.1%\), Brand Memory \(9\.0%\), Outdoor Navigation \(8\.6%\), and Social Chat \(8\.2%\), where relevant evidence must be retrieved and connected across interactions\. The smaller gains on Cartoon Entertainment \(6\.1%\) and Home Renovation \(5\.2%\) are consistent with their greater reliance on direct visual\-semantic recall\. On MemGallery, the largest margins appear in Conflict Detection \(10\.6%\), Knowledge Resolution \(10\.5%\), and Factual Retrieval \(8\.1%\), demonstrating the value of structured access for reconciling and precisely locating information\.AdaMMfurther improves Multi\-entity Reasoning, Visual\-centric Reasoning, and Visual\-centric Search, while remaining comparable to the strongest baselines on Answer Refusal, Temporal Reasoning, and Test\-time Learning\. Together, these results suggest that combining retrieval with operation\-specific analytic access is particularly beneficial when answering requires more than semantic relevance alone\.

![Refer to caption](https://arxiv.org/html/2607.29440v1/x3.png)Figure 3:LLM\-as\-Judge performance across fine\-grained tasks with GPT\-5\.4\-mini\.

### 4\.3Ablation Study

![Refer to caption](https://arxiv.org/html/2607.29440v1/x4.png)Figure 4:Ablation results on MemEye\.Overall and task\-level LLM\-as\-Judge accuracy using GPT\-5\.4\-mini\.We evaluate four ablated variants on MemEye to examine the contributions of the designed memories and the memory\-aware planning mechanism\. The Health task emphasizes numerical and temporal analysis over multiple personal\-health records, whereas the Brand task primarily evaluates visually grounded recall of brand\-related information\. In*w/o Analytic*and*w/o Retrieval*, we disable the analytic memory and retrieval memory, respectively\. In*w/o Planning Context*, the planner receives the entire memory metadata\. In*w/o Progressive Execution*, all tool calls and their arguments are instantiated before execution\.

##### Contributions of Complementary Memories\.

As shown in Figure[4](https://arxiv.org/html/2607.29440#S4.F4), removing either memory component consistently degrades performance\. Without analytic memory, overall accuracy decreases by 4\.6%, with a substantially larger 14\.9% drop on Health\. This result highlights the importance of organizing recurring records into executable structures for analytical queries\. Removing retrieval memory instead causes a 7\.5% overall drop and is particularly detrimental to Brand, where performance decreases by 9\.9%\. The distinct degradation patterns confirm that analytic and retrieval memory provide complementary capabilities\.

##### Contributions of Memory\-Aware Planning\.

Removing the planning context reduces performance by 3\.3% overall and by 6\.7% and 6\.3% on Health and Brand, respectively\. This demonstrates that effective tool selection requires knowledge of both the query intent and the memory structures currently available\. Instantiating all calls before execution also produces consistent degradation, including drops of 4\.2% on Health and 3\.2% on Brand\. These results indicate that progressive execution is useful when later operations depend on information obtained from preceding calls\.

## 5Conclusion

In this work, we identify a retrieval–analysis mismatch in long\-term multimodal agent memory, highlighting the need for executable operations over accumulated experience beyond retrieval\. To address this issue, we introducedAdaMM, which couples hierarchical retrieval memory with schema\-induced analytic memory and exposes their distinct capabilities through operation\-specific tools\. A memory\-aware planner further grounds tool selection in the current memory state and progressively composes retrieval and analytic operations\. Experiments on MemEye and MemGallery across two answer backbones demonstrate consistent improvements from combining flexible recall with executable analysis\.

## Limitations

WhileAdaMMdemonstrates promising results in long\-term multimodal agent memory, two limitations remain\. First, analytic memory depends on record fragments extracted from multimodal interactions\. Incorrect or missing fields may propagate to schema induction, table construction, and downstream computations\. Future work could incorporate confidence\-aware extraction and cross\-round consistency checks to improve robustness\. Second,AdaMMuses a predefined set of access tools and thus requires manual extension for unseen, domain\-specific operations\. Future work could develop a self\-evolving tool framework that identifies emerging capability gaps and safely synthesizes, validates, and integrates new tools\.

## References

- R\. Agrawal and R\. Srikant \(1994\)Fast algorithms for mining association rules in large databases\.InProceedings of the 20th International Conference on Very Large Data Bases,pp\. 487–499\.Cited by:[§3\.3\.2](https://arxiv.org/html/2607.29440#S3.SS3.SSS2.Px1.p1.5)\.
- Y\. Bei, T\. Wei, X\. Ning, Y\. Zhao, Z\. Liu, X\. Lin, Y\. Zhu, H\. Hamann, J\. He, and H\. Tong \(2026\)Mem\-gallery: benchmarking multimodal long\-term conversational memory for mllm agents\.arXiv preprint arXiv:2601\.03515\.Cited by:[§4\.1](https://arxiv.org/html/2607.29440#S4.SS1.SSS0.Px1.p1.1)\.
- 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 arXiv:2504\.19413\.Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px1.p1.1)\.
- D\. Driess, F\. Xia, M\. S\. M\. Sajjadi, C\. Lynch, A\. Chowdhery, B\. Ichter, A\. Wahid, J\. Tompson, Q\. Vuong, T\. Yu, W\. Huang, Y\. Chebotar, P\. Sermanet, D\. Duckworth, S\. Levine, V\. Vanhoucke, K\. Hausman, M\. Toussaint, K\. Greff, A\. Zeng, I\. Mordatch, and P\. Florence \(2023\)PaLM\-e: an embodied multimodal language model\.arXiv preprint arXiv:2303\.03378\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1)\.
- J\. Feng, B\. Xu, J\. Chen, M\. Dai, C\. Wu, H\. Li, B\. Zeng, Y\. Xie, H\. Liang, M\. Lu, and W\. Zhang \(2026\)M2A: multimodal memory agent with dual\-layer hybrid memory for long\-term personalized interactions\.arXiv preprint arXiv:2602\.07624\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1),[§1](https://arxiv.org/html/2607.29440#S1.p2.1),[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2607.29440#S4.SS1.SSS0.Px2.p1.1)\.
- M\. Guo, Q\. Jiao, Z\. Shi, Y\. Quan, B\. Zhang, D\. Li, L\. Che, W\. Xu, S\. Liu, Z\. Liu, M\. Kapadia, V\. Pavlovic, J\. Liu, M\. Wang, Y\. Shi, D\. N\. Metaxas, and R\. Tang \(2026\)MemEye: a visual\-centric evaluation framework for multimodal agent memory\.arXiv preprint arXiv:2605\.15128\.Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2607.29440#S4.SS1.SSS0.Px1.p1.1)\.
- K\. Hatalis, D\. Christou, J\. Myers, S\. Jones, K\. Lambert, A\. Amos\-Binks, Z\. Dannenhauer, and D\. Dannenhauer \(2023\)Memory matters: the need to improve long\-term memory in llm\-agents\.InProceedings of the AAAI Symposium Series,Vol\.2,pp\. 277–280\.Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px1.p1.1)\.
- H\. He, W\. Yao, K\. Ma, W\. Yu, Y\. Dai, H\. Zhang, Z\. Lan, and D\. Yu \(2024\)WebVoyager: building an end\-to\-end web agent with large multimodal models\.arXiv preprint arXiv:2401\.13919\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1)\.
- Z\. Huang, Z\. Tian, Q\. Guo, F\. Zhang, Y\. Zhou, D\. Jiang, Z\. Xie, and X\. Zhou \(2025\)LiCoMemory: lightweight and cognitive agentic memory for efficient long\-term reasoning\.arXiv preprint arXiv:2511\.01448\.Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px1.p1.1)\.
- J\. Y\. Koh, R\. Lo, L\. Jang, V\. Duvvur, M\. C\. Lim, P\. Huang, G\. Neubig, S\. Zhou, R\. Salakhutdinov, and D\. Fried \(2024\)VisualWebArena: evaluating multimodal agents on realistic visual web tasks\.arXiv preprint arXiv:2401\.13649\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1)\.
- Y\. Li, H\. Wen, W\. Wang, X\. Li, Y\. Yuan, G\. Liu, J\. Liu, W\. Xu, X\. Wang, Y\. Sun,et al\.\(2024\)Personal llm agents: insights and survey about the capability, efficiency and security\.arXiv preprint arXiv:2401\.05459\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1),[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px1.p1.1)\.
- Z\. Li, S\. Song, H\. Wang, S\. Niu, D\. Chen, J\. Yang, C\. Xi, H\. Lai, J\. Zhao, Y\. Wang, J\. Ren, Z\. Lin, J\. Huo, T\. Chen, K\. Chen, K\. Li, Z\. Yin, Q\. Yu, B\. Tang, H\. Yang, Z\. J\. Xu, and F\. Xiong \(2025\)MemOS: an operating system for memory\-augmented generation \(MAG\) in large language models\.CoRRabs/2505\.22101\.Cited by:[§4\.1](https://arxiv.org/html/2607.29440#S4.SS1.SSS0.Px2.p1.1)\.
- J\. Liu, Z\. Ling, S\. Qiu, Y\. Liu, S\. Han, P\. Xia, H\. Tu, Z\. Zheng, C\. Xie, C\. Fleming, M\. Ding, and H\. Yao \(2026\)Omni\-simplemem: autoresearch\-guided discovery of lifelong multimodal agent memory\.arXiv preprint arXiv:2604\.01007\.Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Lu, W\. Cheng, Z\. Zhang, and H\. Tang \(2026\)MMA: multimodal memory agent\.arXiv preprint arXiv:2602\.16493\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p2.1),[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2607.29440#S4.SS1.SSS0.Px2.p1.1)\.
- V\. Nguyen, T\. Nguyen, V\. M\. Patel, and Y\. Li \(2026\)Personal visual memory from explicit and implicit evidence\.arXiv preprint arXiv:2605\.28806\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p2.1),[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px2.p1.1)\.
- E\. R\. Omiecinski \(2003\)Alternative interest measures for mining associations in databases\.IEEE Transactions on Knowledge and Data Engineering15\(1\),pp\. 57–69\.External Links:[Document](https://dx.doi.org/10.1109/TKDE.2003.1161582)Cited by:[§3\.3\.2](https://arxiv.org/html/2607.29440#S3.SS3.SSS2.Px2.p1.3)\.
- 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 arXiv:2310\.08560\.Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px1.p1.1)\.
- A\. Petrov, A\. Gusak, D\. Mukha, and D\. Korolev \(2026\)From unstructured recall to schema\-grounded memory: reliable AI memory via iterative, schema\-aware extraction\.arXiv preprint arXiv:2604\.27906\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2604.27906)Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px1.p1.1)\.
- \[19\]P\. Rasmussen, P\. Paliychuk, T\. Beauvais, J\. Ryan, and D\. ChalefZep: a temporal knowledge graph architecture for agent memory, 2025\.URL https://arxiv\. org/abs/2501\.13956\.Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px1.p1.1)\.
- T\. Schick, J\. Dwivedi\-Yu, R\. Dessì, R\. Raileanu, M\. Lomeli, L\. Zettlemoyer, N\. Cancedda, and T\. Scialom \(2023\)Toolformer: language models can teach themselves to use tools\.arXiv preprint arXiv:2302\.04761\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1)\.
- G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. Anandkumar \(2023\)Voyager: an open\-ended embodied agent with large language models\.arXiv preprint arXiv:2305\.16291\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1)\.
- X\. Wang, Y\. Zhou, X\. Liu, H\. Lu, Y\. Xu, F\. He, J\. Yoon, T\. Lu, G\. Bertasius, M\. Bansal, H\. Yao, and F\. Huang \(2024\)Mementos: a comprehensive benchmark for multimodal large language model reasoning over image sequences\.arXiv preprint arXiv:2401\.10529\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1)\.
- Y\. Wang and X\. Chen \(2025\)Mirix: multi\-agent memory system for llm\-based agents\.arXiv preprint arXiv:2507\.07957\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1),[§1](https://arxiv.org/html/2607.29440#S1.p2.1),[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2607.29440#S4.SS1.SSS0.Px2.p1.1)\.
- T\. Xie, D\. Zhang, J\. Chen, X\. Li, S\. Zhao, R\. Cao, T\. J\. Hua, Z\. Cheng, D\. Shin, F\. Lei, Y\. Liu, Y\. Xu, S\. Zhou, S\. Savarese, C\. Xiong, V\. Zhong, and T\. Yu \(2024\)OSWorld: benchmarking multimodal agents for open\-ended tasks in real computer environments\.arXiv preprint arXiv:2404\.07972\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1)\.
- W\. Xu, K\. Mei, H\. Gao, J\. Tan, Z\. Liang, and Y\. Zhang \(2025\)A\-mem: agentic memory for llm agents\.arXiv preprint arXiv:2502\.12110\.Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px1.p1.1),[§4\.1](https://arxiv.org/html/2607.29440#S4.SS1.SSS0.Px2.p1.1)\.
- S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao \(2022\)ReAct: synergizing reasoning and acting in language models\.arXiv preprint arXiv:2210\.03629\.Cited by:[§1](https://arxiv.org/html/2607.29440#S1.p1.1)\.
- J\. Yue, C\. Hu, J\. Sheng, Z\. Zhou, W\. Zhang, T\. Liu, L\. Guo, and Y\. Deng \(2026\)Hypermem: hypergraph memory for long\-term conversations\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 35237–35254\.Cited by:[§3\.4](https://arxiv.org/html/2607.29440#S3.SS4.p2.2)\.
- Z\. Zhang, Q\. Dai, X\. Bo, C\. Ma, R\. Li, X\. Chen, J\. Zhu, Z\. Dong, and J\. Wen \(2025\)A survey on the memory mechanism of large language model\-based agents\.ACM Transactions on Information Systems43\(6\),pp\. 1–47\.Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px1.p1.1)\.
- W\. Zhong, L\. Guo, Q\. Gao, H\. Ye, and Y\. Wang \(2024\)MemoryBank: enhancing large language models with long\-term memory\.InAAAI,pp\. 19724–19731\.Cited by:[§2](https://arxiv.org/html/2607.29440#S2.SS0.SSS0.Px1.p1.1)\.