Paper deep dive

Anatomy of Agentic Memory: Taxonomy and Empirical Analysis of Evaluation and System Limitations

Dongming Jiang, Yi Li, Songtao Wei, Jinxin Yang, Ayushi Kishore, Alysa Zhao, Dingyi Kang, Xu Hu, Feng Chen, Qiannan Li, Bingzhe Li

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 83

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%

Last extracted: 7/20/2026, 8:09:53 PM

Summary

This survey analyzes Memory-Augmented Generation (MAG) systems, proposing a taxonomy of four memory structures (Lightweight Semantic, Entity-Centric/Personalized, Episodic/Reflective, and Structured/Hierarchical) and identifying key empirical limitations such as benchmark saturation, metric misalignment, backbone sensitivity, and system-level latency/throughput costs.

Entities (11)

GPT-4o-mini · model · 98%Memory-Augmented Generation · concept · 98%Qwen-2.5-3B · model · 98%Entity-Centric Memory · memorystructure · 95%Episodic Memory · memorystructure · 95%HotPotQA · benchmark · 95%Lightweight Semantic Memory · memorystructure · 95%LongMemEval-M · benchmark · 95%Structured Memory · memorystructure · 95%Benchmark Saturation · limitation · 92%Latency Overhead · limitation · 90%

Relation Signals (12)

Qwen-2.5-3B → usedin → Memory-Augmented Generation

confidence 98% · We employ a suite of Large Language Models (LLMs) to serve as the agent controller including ... Qwen-2.5-3B

GPT-4o-mini → usedin → Memory-Augmented Generation

confidence 98% · We employ a suite of Large Language Models (LLMs) to serve as the agent controller including gpt-4o-mini

LongMemEval-M → hassaturationrisk → Low

confidence 95% · Only datasets that substantially exceed the active window (e.g., LongMemEval-M at >1M tokens) structurally require external memory

HotPotQA → hassaturationrisk → High

confidence 95% · Benchmarks such as HotpotQA (~1k tokens) ... fall within a 128k context window, implying high theoretical saturation risk

Memory-Augmented Generation → hassubcategory → Lightweight Semantic Memory

confidence 95% · We introduce a concise taxonomy of Memory-Augmented Generation organized around four memory structures: Lightweight Semantic...

Memory-Augmented Generation → hassubcategory → Entity-Centric Memory

confidence 95% · We introduce a concise taxonomy of Memory-Augmented Generation organized around four memory structures: ... Entity-Centric and Personalized...

Memory-Augmented Generation → hassubcategory → Episodic Memory

confidence 95% · We introduce a concise taxonomy of Memory-Augmented Generation organized around four memory structures: ... Episodic and Reflective...

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Agentic memory systems enable large language model (LLM) agents to maintain state across long interactions, supporting long-horizon reasoning and personalization beyond fixed context windows. Despite rapid architectural development, the empirical foundations of these systems remain fragile: existing benchmarks are often underscaled, evaluation metrics are misaligned with semantic utility, performance varies significantly across backbone models, and system-level costs are frequently overlooked. This survey presents a structured analysis of agentic memory from both architectural and system perspectives. We first introduce a concise taxonomy of MAG systems based on four memory structures. Then, we analyze key pain points limiting current systems, including benchmark saturation effects, metric validity and judge sensitivity, backbone-dependent accuracy, and the latency and throughput overhead introduced by memory maintenance. By connecting the memory structure to empirical limitations, this survey clarifies why current agentic memory systems often underperform their theoretical promise and outlines directions for more reliable evaluation and scalable system design.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

81,418 characters extracted from source content.

Expand or collapse full text

Anatomy of Agentic Memory: Taxonomy and Empirical Analysis of Evaluation and System Limitations Dongming Jiang α , Yi Li α , Songtao Wei α , Jinxin Yang α , Ayushi Kishore β , Alysa Zhao γ , Dingyi Kang α , Xu Hu α , Feng Chen α , Qiannan Li β and Bingzhe Li α,∗ α University of Texas at Dallas β University of California, Davis γ Texas A&M University dongming.jiang, yi.li3, songtao.wei, jinxin.yang, dingyi.kang, xu.hu, feng.chen, bingzhe.li@utdallas.edu aykishore, qnli@ucdavis.edu; alysazhao111@tamu.edu ∗ Corresponding author§ GitHub Repo Abstract Agentic memory systems enable large language model (LLM) agents to maintain state across long interactions, supporting long-horizon rea- soning and personalization beyond fixed con- text windows. Despite rapid architectural de- velopment, the empirical foundations of these systems remain fragile: existing benchmarks are often underscaled, evaluation metrics are misaligned with semantic utility, performance varies significantly across backbone models, and system-level costs are frequently over- looked. This survey presents a structured anal- ysis of agentic memory from both architectural and system perspectives. We first introduce a concise taxonomy of Memory-Augmented Generation (MAG) systems based on four memory structures. Then, we analyze key pain points limiting current systems, includ- ing benchmark saturation effects, metric valid- ity and judge sensitivity, backbone-dependent accuracy, and the latency and throughput over- head introduced by memory maintenance. By connecting the memory structure to empirical limitations, this survey clarifies why current agentic memory systems often underperform their theoretical promise and outlines directions for more reliable evaluation and scalable sys- tem design. To facilitate future research, we maintain an open-source repository of surveyed papers, benchmarks, and resources. 1 1 Introduction Large language model (LLM) agents are increas- ingly expected to operate over long time hori- zons, maintaining user preferences, accumulating task-relevant knowledge, etc. (Brown et al., 2020; Achiam et al., 2023; Wei et al., 2022). How- ever, fixed context windows fundamentally limit 1 https://github.com/FredJiang0324/ Anatomy-of-Agentic-Memory.Given the rapid evo- lution of research in agentic memory, we may have inadvertently overlooked some relevant studies. We welcome you to suggest missing papers by emailing us or opening an issue on our GitHub repository. their ability to retain and manipulate persistent state (Brown et al., 2020; Beltagy et al., 2020; Liu et al., 2024; Press et al., 2021). To address this constraint, Memory-Augmented Generation (MAG) extends memory beyond the prompt, en- abling agents to store, retrieve, and update infor- mation across interactions (Xu et al., 2025c; Nan et al., 2025; Chhikara et al., 2025; Jiang et al., 2026a; Liu et al., 2026). While this paradigm has rapidly evolved from lightweight semantic stores to entity-centric, reflective, and hierarchical designs, empirical understanding remains limited: reported gains are inconsistent across benchmarks, highly backbone-dependent, and lack principled guidance on evaluation and system-level cost. These challenges stem in part from underspec- ified design trade-offs and inadequate evaluation. Benchmarks are often underscaled relative to mod- ern context windows, metrics emphasize surface overlap over semantic utility, and system-level costs such as latency and throughput degradation are rarely measured. As a result, complex mem- ory systems are frequently tested in settings where simpler full-context or retrieval baselines suffice, obscuring their true benefits and limitations. In this paper, we provide a structured analysis of agentic memory systems from both architec- tural and empirical perspectives. 1) We introduce a concise taxonomy of Memory-Augmented Gen- eration organized around four memory structures: Lightweight Semantic, Entity-Centric and Person- alized, Episodic and Reflective, and Structured and Hierarchical. Defined by how memory is orga- nized and manipulated, this taxonomy establishes a principled basis for analyzing system behavior. 2) Building on this framework, we identify key bottle- necks limiting reliability and scalability, including benchmark saturation, metric misalignment (e.g., F1 vs. semantic correctness), prompt sensitivity in LLM-as-a-judge evaluation, backbone dependence, and system-level costs such as retrieval latency, arXiv:2602.19320v2 [cs.CL] 20 May 2026 Table 1: Comparison with related surveys on memory for LLM-based agents.✓indicates the topic is systematically discussed; (✓) indicates partial or brief coverage; × indicates the topic is not addressed. SurveyTaxonomy Focus Memory Mgmt. & Policy Benchmark Saturation Metric Validity Backbone Sensitivity System Cost & Latency The AI Hippocampus (Jia et al., 2026)Brain-inspired: implicit, explicit, agentic(✓)× Memory in the Age of AI Agents (Hu et al., 2025)Forms–functions–dynamics(✓)× Toward Efficient Agents (Yang et al., 2026c)Efficiency-focused: memory, tool learning, planning✓×✓ Rethinking Memory Mechanisms (Huang et al., 2026)Substrate–cognition–subject✓(✓)× From Storage to Experience (Luo et al., 2026)Evolutionary: storage–reflection–experience(✓)× Graph-based Agent Memory (Yang et al., 2026a)Graph-oriented lifecycle✓(✓)× Taxonomy and Empirical Analysis (Ours)Structural + Empirical analysis✓ update overhead, and throughput degradation. By linking memory structures to empirical lim- itations, this survey clarifies why current agentic memory systems often fall short of their theoretical promise. Rather than advocating a single “best” design, we provide a diagnostic framework to ex- plain when specific memory structures are effective, when they fail, and what trade-offs they entail. Our analysis offers guidance for designing more robust benchmarks, more reliable evaluation protocols, and more scalable agentic memory systems. Difference from other surveys: While existing surveys (Jia et al., 2026; Hu et al., 2025; Yang et al., 2026c,a; Luo et al., 2026; Huang et al., 2026) pri- marily operate at the theoretical level by catalogu- ing architectures, defining conceptual taxonomies, and drawing cognitive science analogies, our sur- vey bridges the gap from theory to practice. Our taxonomy is structure-oriented, not only discussing various memory structure designs, but also high- lighting the memory management frameworks and optimization strategies. In addition, we provide comprehensive evaluations across multiple bench- marks. Specifically, we conduct systematic analy- ses of agentic memory systems on benchmark sat- uration, metric validity, backbone sensitivity, and maintenance overhead, overlooked in prior surveys yet critical for understanding why current MAG systems often fall short of their theoretical promise. A detailed comparison is presented in Table 1. 2 Background Agentic memory extends retrieval-based genera- tion by introducing persistent, writable memory that evolves across interactions, enabling an LLM agent to store, update, and reuse information over time. Formally, at stept, the agent conditions on observationso t and an external memory stateM t : y t ∼ f θ φ(o t ,s t ) ⊕ ψ(M t ;q t ) ,(1) wherey t denotes the output,s t additional agent state, andψ(M t ;q t )retrieves memory given query q t . The operator⊕represents integration (e.g., prompt concatenation or structured fusion). Cru- cially, memory affects behavior through the explicit retrieval term ψ(M t ;q t ) rather than updates to θ. Two coupled processes are operated in agentic memory: inference-time recall and memory up- date. At each step, the agent retrieves relevant information from an external memory store to con- dition its decision, and subsequently writes, up- dates, or consolidates memory to maintain a useful long-term state. Unlike parametric learning, this mechanism influences behavior through explicit read–write operations over an evolving memory state rather than by modifying model weights. A formalization of these operations including query generation, utility-aware retrieval, and memory ac- tions such as store, summarize, link, and delete is provided in Appendix B. 3 Taxonomy of Agentic Memory We introduce a concise taxonomy of Memory- Augmented Generation organized around four memory structures: Lightweight Semantic, Entity- Centric and Personalized, Episodic and Reflective, and Structured and Hierarchical. Each category is further split into subcategories as shown in Figure 1 in Appendix A. 3.1 Lightweight Semantic Memory Lightweight Semantic Memory is the simplest and most widely used form of MAG, where memory consists of independent textual units embedded in a vector space and retrieved via top-k similarity search. Entries are typically append-only or mini- mally filtered, with no explicit structural relations between them. RL-Optimized Semantic Compression: These schemes treat memory as a fixed-size semantic store and apply RL to optimize how information is compressed, retained, or overwritten under context constraints (Wang et al., 2025b; Yan et al., 2025; Yuan et al., 2025). Memory remains largely un- structured and textual, with learning focused on efficient content selection. For example, MemA- gent (Yu et al., 2025) trains a latent token-level memory using multi-conversation RL to manage ultra-long contexts, while MemSearcher (Yuan et al., 2025) formulates multi-turn search and mem- ory updates as an end-to-end RL problem, iter- atively compressing semantic memory to enable scalable multi-hop reasoning without relying on full dialogue history. Heuristic / Prompt-Optimized: These approaches manage memory through prompt design or heuris- tic rewriting with a flat, compressed textual sum- mary of prior steps generated via engineered in- structions, reducing context length but remaining unstructured. Similar prompt-driven compression strategies appear in prior work (Zhao et al., 2025; Wu and Li, 2025; Liu et al., 2026; Li et al., 2026b). For example, ACON (Kang et al., 2025c) learns natural-language compression guidelines to selec- tively summarize long interaction histories, reduc- ing context by up to 54% without RL or fine-tuning, while CISM (Liu et al., 2025c) condenses each reasoning and action step into compact semantic representations to enable long-horizon execution under context constraints without explicit external memory retrieval. Context Window Management: This category manages the model’s working context within a sin- gle task, without accumulating memory across ses- sions. Prior interactions are folded, summarized, or reorganized to fit within a bounded window, pri- oritizing local reasoning efficiency over long-term storage or reuse (Zhu et al., 2025; Sakib et al., 2025). For example, AgentFold (Ye et al., 2025) treats context as a dynamic workspace and learns multi-scale folding operations to condense long tra- jectories, while Context-Folding Agent (Sun et al., 2025b) trains an RL-based policy that branches sub-tasks and compresses completed segments. Token-Level Semantic Memory: This category encodes memory at the token level using dedicated memory tokens or compressed latent panels. These representations primarily capture semantic content, aiming to improve long-context handling with min- imal overhead (Wu et al., 2025b; Zhang et al., 2025c; Yang et al., 2024b). Memory entries are independent and inexpensive to store or retrieve, making them suitable for short- to medium-horizon recall, but limited for precise state tracking or long- term reasoning. For example, MemGen (Zhang et al., 2025c) augments a frozen LLM with on- demand latent token memory via an RL-trained trig- ger and LoRA-based weaving, while TokMem (Wu et al., 2025b) replaces lengthy procedural prompts with trainable memory tokens to enable constant- size context management and scalable skill reuse. 3.2 Entity-Centric and Personalized Memory Entity-centric and personalized memory organizes information around explicit entities such as users, tasks, or preferences, using structured records or attribute–value pairs. A predefined schema governs how information is stored, updated, and retrieved. Entity-Centric Memory: Entity-centric mem- ory organizes information around explicit entities and their attributes, maintaining structured, persis- tent records rather than raw dialogue (Modarressi et al., 2023; Liu et al., 2021, 2025a). For exam- ple, A-MEM (Xu et al., 2025c) builds intercon- nected knowledge notes with structured attributes and LLM-generated links; Memory-R1 (Yan et al., 2025) formulates entity memory management as an RL problem over a persistent entity–fact bank. Personalized Memory: Personalized memory maintains persistent user profiles that integrate short- and long-term preferences to support adaptive, identity-consistent behavior across ses- sions (Zhong et al., 2024; Li et al., 2025a; Kwon et al., 2025; Liu et al., 2025a; Mao et al., 2026; Su et al., 2026).For example, PAMU (Sun et al., 2025a) combines sliding windows with moving averages to track evolving preferences, EgoMem (Yao et al., 2025) constructs lifelong mul- timodal profiles with conflict-aware updates, and MemOrb (Huang et al., 2025a) stores compact re- flective memories for continual improvement. 3.3 Episodic and Reflective Memory Episodic and reflective memory adds temporal ab- straction by organizing interactions into episodes or higher-level summaries. These systems period- ically consolidate experience through summariza- tion or reflection, producing compact representa- tions of salient events over time. Episodic Buffer w/ Learned Control: mem- ory in these work consists of episodic interaction records maintained in a bounded buffer and dynam- ically inserted, retained, or deleted through learned policies (Du et al., 2025a; Zhang et al., 2025f; Icarte et al., 2020). For example, MemR 3 (Du et al., 2025a) models retrieval as a closed-loop re- trieve–reflect–answer process; and the Act of Re- membering (Icarte et al., 2020) formulates remem- bering as a control problem in POMDPs with a fixed-capacity episodic buffer. Episodic Recall for Exploration: These meth- ods leverage episodic memory to improve explo- ration and credit assignment in partially observ- able or long-horizon settings. Past experiences are stored and selectively retrieved to guide decision- making (Na et al., 2024; Adamyan et al., 2025). For example, EMU (Na et al., 2024) maintains large- capacity episodic memories indexed by learned embeddings to accelerate cooperative MARL ex- ploration, while SAM2RL (Adamyan et al., 2025) uses a visual memory bank as an episodic buffer and trains a PPO policy to manage memory re- placement, outperforming heuristic updates under challenging conditions. Episodic Reflection & Consolidation: This sub- category reflects and consolidates episodic expe- riences into compact representations (Tan et al., 2025b; Kim et al., 2025; Ouyang et al., 2025; Dong et al., 2025; Lee et al., 2024). The ob- jective is to balance memory capacity with long- term reasoning utility. For example, MemP (Fang et al., 2025b) distills trajectories into procedural abstractions for continual refinement and transfer; LEGOMem (Han et al., 2025) constructs modular, role-aware procedural memories for multi-agent co- ordination; and TiMem (Li et al., 2026a) introduces a temporal-hierarchical memory tree for structured consolidation and scalable long-horizon personal- ization without RL or fine-tuning. Episodic Utility Learning: Episodic memories in these setting are augmented with learned value or utility signals that evolve over time, enabling selective retention and retrieval based on both se- mantic relevance and estimated long-term useful- ness (Zhou et al., 2025a; Cao et al., 2025). For example, MemRL (Zhang et al., 2026c) associates utility Q-values with intent–experience pairs and updates them online to balance stability and plas- ticity without fine-tuning, while Memory-T1 (Du et al., 2025b) learns a temporal-aware retrieval pol- icy via GRPO to optimize accuracy, grounding, and chronological consistency in long-context dia- logue. 3.4 Structured and Hierarchical Memory Structured and hierarchical memory systems im- pose explicit organization over stored information. Hierarchical designs partition memory into multi- ple tiers (e.g., short- and long-term stores), while structured approaches encode relationships among memory elements using graphs or other formal re- lational representations. Graph-Structured Memory: Graph-structured memory represents information as nodes and edges capturing semantic, temporal, causal, or entity- level relations, enabling reasoning over structured subgraphs (Zhang et al., 2025b,d; Jiang et al., 2026c; Tao et al., 2026; Zhang et al., 2026d; Hu et al., 2026b). This design supports multi-hop in- ference, provenance tracking, and coherent long- horizon reasoning. For example, MAGMA (Jiang et al., 2026a) organizes memory across semantic, temporal, causal, and entity graphs; Zep (Ras- mussen et al., 2025) constructs a bi-temporal knowledge graph with episodic and semantic lay- ers; SGMem (Wu et al., 2025a) models dia- logue as sentence-level graphs; and LatentGraph- Mem (Zhang et al., 2026d) integrates latent graph encoding with a compact symbolic subgraph to balance stability, efficiency, and interpretability. OS-Inspired & Hierarchical Memory: OS- inspired and hierarchical memory systems orga- nize information into multi-tier storage layers (e.g., short-term, episodic, long-term), dynamically mov- ing and consolidating data to balance scalability, re- tention, and adaptive forgetting (Xu, 2025; Ouyang, 2025; Zhang et al., 2025e; Jia et al., 2025; Li et al., 2026a). For example, MemGPT (Packer et al., 2023) enables LLM-driven memory paging across tiers; MemoryOS (Kang et al., 2025a) implements a modular three-level hierarchy; EverMemOS (Hu et al., 2026a) and HiMem (Zhang et al., 2026b) consolidate episodic and semantic traces for long- horizon adaptation; and MeMAD (Ling et al., 2025) stores structured debate experiences for reusable reasoning. Policy-Optimized Memory Management: Policy- optimized memory management treats storage, up- date, consolidation, and deletion as learnable de- cisions, using reinforcement learning or hybrid training to optimize long-horizon rewards (Liu et al., 2025b; Xu et al., 2025b; Kang et al., 2025b; Du et al., 2025b). For example, MEM1 (Zhou et al., 2025b) learns to maintain a compact inter- nal state with constant-memory operations; and Mem-α(Wang et al., 2025b) trains an RL policy to manage multi-component external memory un- der ultra-long contexts; and AtomMem (Huo et al., 2026) decomposes memory into CRUD actions to learn task-aligned control strategies. While en- abling adaptive and scalable management, these approaches introduce greater system complexity and nontrivial maintenance overhead. 3.5 Discussion The four categories described above capture the dominant memory structures used in contemporary MAG systems. While individual systems may com- bine multiple mechanisms, each can typically be characterized by a primary memory organization that governs its behavior. This structure-first taxon- omy provides a foundation for understanding how design choices in agentic memory influence accu- racy, efficiency, and reliability. In the next section, we build on this taxonomy to analyze the empirical limitations and pain points that arise across current MAG systems. 4 Evaluation and Pain Points In this section, we move beyond taxonomy to em- pirically analyze the practical bottlenecks hinder- ing robust deployment. While theoretical architec- tures are promising, real world utility is strictly con- strained by evaluation validity, system efficiency, and backbone reliability. We dissect these chal- lenges across four critical dimensions: 1.Benchmark Validity: Are we testing memory or just context length? 2.Metric Reliability: Can lexical metrics capture semantic coherence? 3. System Efficiency: The “Agency Tax” of latency and cost. 4.Backbone Sensitivity: The “Silent Failure” of memory operations in open-weight models. 4.1 Experimental Setup We evaluate representative MAG systems span- ning the four taxonomy categories introduced in Section 3. Six memory architectures are selected: AMem (Xu et al., 2025c), MemoryOS (Kang et al., 2025a), Nemori (Nan et al., 2025), MAGMA (Jiang et al., 2026a), SimpleMEM (Liu et al., 2026) and MemSkill (Zhang et al., 2026a) as shown in Table 8 of Appendix E. All systems are configured to fol- low their default or recommended settings, except where modifications are required to ensure com- parability. We employ a suite of Large Language Models (LLMs) to serve as the agent controller includinggpt-4o-mini(Achiam et al., 2023) and Qwen-2.5-3B (Yang et al., 2024a). 4.2 Benchmark Scalability: The Context Saturation Risk A key motivation for agentic memory is to support reasoning beyond a model’s finite context window. Yet as LLM windows expand (e.g., 128k to 1M tokens), many benchmarks risk context saturation, where all relevant information fits within a single prompt, making external memory seemingly un- necessary. In this section, rather than comparing performance, we examine the intrinsic properties of existing datasets to evaluate their continued validity in the long-context era. 4.2.1 Dimensions of Limitation First, we evaluate benchmark scalability along three structural axes: volume, interaction depth, and entity diversity, to assess their saturation risk under long-context LLMs as shown in Table 2. Volume (Total Token Load). This dimension captures the aggregate information size a model must process. Benchmarks such as HotpotQA (∼1k tokens) and MemBench (∼100k tokens) fall within a 128k context window, implying high theoretical saturation risk. LoCoMo (∼20k tokens) similarly remains comfortably in-window for modern mod- els. Only datasets that substantially exceed the active window (e.g., LongMemEval-M at >1M to- kens) structurally require external memory. Interaction Depth (Temporal Structure). Be- yond raw volume, scalability depends on how in- formation unfolds across sessions. Single-turn QA (e.g., HotpotQA) imposes minimal temporal de- pendency, whereas multi-session settings (e.g., Lo- CoMo with 35 sessions) introduce longitudinal rea- soning. However, unless cross-session dependen- cies exceed the active window or require persis- tent state tracking beyond prompt capacity, such datasets may still be solvable through direct in- context aggregation rather than true memory man- agement. Entity Diversity (Relational Complexity). This axis measures how many distinct entities or conceptual threads must be tracked simultaneously. Low-diversity benchmarks permit near-isolated re- Table 2: Structural saturation risk of memory benchmarks. Benchmarks are analyzed based on intrinsic statistics rather than model performance. Saturation Risk is a heuristic estimate of whether a long-context LLM may solve the benchmark through direct prompting. Benchmark Scalability Dimensions Theoretical Saturation Risk Avg. VolumeInteraction DepthEntity Diversity HotpotQA (Yang et al., 2018)∼1k TokensSingle TurnLowHigh (Trivial for Context Window) LoCoMo (Maharana et al., 2024)∼20k Tokens35 SessionsHighModerate (Requires Reasoning) LongMemEval-S (Wu et al., 2024)103k Tokens5 Core AbilitiesHighModerate (Borderline) LongMemEval-M (Wu et al., 2024) >1M Tokens5 Core AbilitiesHighLow (Requires External Memory) MemBench (Tan et al., 2025a)∼100k TokensFact/ReflectionMediumHigh (Fits in 128k Window) trieval, while higher-diversity settings (e.g., Lo- CoMo, LongMemEval) increase interference and relational reasoning demands. Nevertheless, if en- tity interactions remain bounded within context lim- its, structured external memory may not be strictly necessary. Discussion: Taken together, these dimensions show that saturation risk is determined not by sur- face difficulty but by whether a benchmark’s struc- tural properties exceed the representational capac- ity of long-context LLMs. 4.2.2Context Saturation Gap as an Empirical Diagnostic To address these limitations, we propose that future evaluations explicitly quantify the Context Satura- tion Gap (∆), defined as the performance differ- ence between a Memory-Augmented Agent (MAG) and a brute-force Full-Context baseline under the same backbone and evaluation protocol: ∆ = Score MAG − Score FullContext (2) A large positive∆indicates that external mem- ory provides an advantage beyond simply placing all available evidence in the prompt, especially in out-of-window or lost-in-the-middle regimes. How- ever,∆should be viewed as a diagnostic rather than a strict pass/fail criterion, since benchmarks may still evaluate memory through efficiency, update- ability, robustness, or evidence faithfulness even when Full-Context remains competitive. Table 2 therefore reports structural satura- tion risk rather than an empirical saturation test. Datasets with limited volume and shallow complex- ity should be paired with Full-Context baselines before drawing strong conclusions about memory- specific benefits. 4.3 LLM-as-a-Judge Evaluation Traditional lexical metrics (e.g., F1, BLEU) em- phasize surface-level token overlap, which is in- sufficient for agentic memory tasks where the goal is accurate retrieval and coherent synthesis rather than exact phrasing. To better capture semantic cor- rectness, we adopt an LLM-based evaluator (gpt- 4o-mini) as a proxy for human judgment. In this section, we assess the reliability of this protocol by analyzing the misalignment between lexical and semantic metrics and demonstrating the stability of our system rankings across competitive evaluation settings. 4.3.1 The Misalignment Gap Do lexical metrics correctly identify the best mem- ory system? To examine this, we compared system rankings produced by F1-score with those gener- ated by an LLM-based judge across six representa- tive architectures on the LoCoMo dataset. Table 3 (Left) reveals a significant disconnect. Lexical metrics often fail to capture the strengths of abstractive memory systems. For example, AMem achieves solid semantic performance (Rank 4 across prompts) due to its logical coherence, yet it is heavily penalized by F1 (Rank 5, Score 0.116) because it does not rely on verbatim overlap. In contrast, SimpleMem receives a relatively higher F1 score (0.268) despite demonstrating limited abil- ity to synthesize complex answers (semantic score < 0.30). This divergence indicates that optimizing solely for F1 may favor surface-level memorization over genuine reasoning and memory integration. 4.3.2 Semantic Judge Robustness Across Prompts A common concern with LLM-as-a-judge is “prompt overfitting,” where a system appears strong only under a specific grading instruction. To ensure fairness and generality, we evaluated all architec- tures using three distinct prompt protocols derived from different sources (details in Appendix D.3). As shown in Table 3 (Right), compared with F1- based rankings, the semantic judge exhibits strong robustness: the relative ordering of architectures remains highly consistent across different rubrics. Table 3: Robustness of system ranking across evaluation protocols. We compare Lexical metrics (F1) against LLM-based semantic evaluation using three distinct prompt sources: MAGMA, Nemori, and SimpleMem. Method Lexical MetricSemantic Judge Score (Rank) F1-ScoreRankPrompt 1 (MAGMA)Prompt 2 (Nemori)Prompt 3 (SimpleMem) AMem (Xu et al., 2025c)0.11650.480 (4)0.512 (4)0.482 (4) MemoryOS (Kang et al., 2025a)0.41330.553 (3)0.589 (3)0.552 (3) Nemori (Nan et al., 2025)0.50210.602 (2)0.781 (1)0.649 (2) MAGMA (Jiang et al., 2026a) 0.46720.670 (1)0.741 (2)0.665 (1) SimpleMEM (Liu et al., 2026)0.26840.294 (5)0.298 (5)0.289 (5) MemSkill (Zhang et al., 2026a)0.08260.221 (6)0.220 (6)0.127 (6) While absolute scores fluctuate due to variations in grading strictness and prompt formulation, the comparative conclusions remain stable. 4.3.3 Discussion Lexical metrics provide a convenient baseline but systematically diverge from semantic judgments due to two core failure modes: the Paraphrase Penalty, where correct abstractive answers are pe- nalized for low token overlap, and the Negation Trap, where high overlap masks factual errors. De- tailed examples are provided in Appendix F. In contrast, the semantic judge demonstrates greater stability: architecture rankings remain con- sistent across different grading rubrics, suggesting it better reflects underlying memory quality rather than surface phrasing. Although absolute scores vary with prompt strictness and some models show rubric-aligned specialization, the relative ordering is robust. Overall, these results support LLM-as-a-judge as a more reliable evaluation protocol for agentic memory, while highlighting the importance of care- ful prompt design. 4.4Backbone Sensitivity and Format Stability Agentic memory requires the backbone model to both answer queries and execute structured mem- ory operations (e.g., updates and consolidation). Long-term stability thus depends on reliable adher- ence to strict output formats. To evaluate this “Sta- bility Gap,” we compare representative memory architectures using an API model (gpt-4o-mini) and an open-weight model (Qwen-2.5-3B). Table 4 reveals a clear divergence driven by in- valid structured outputs (e.g., malformed JSON, hallucinated keys) during memory maintenance: 1) Instruction Following vs. Reasoning: While Qwen-2.5-3Bdemonstrates basic capability in con- versational reasoning, it experiences a noticeable drop in End-Task Answer Scores and exhibits a sig- nificantly higher format error rate during memory Table 4: Backbone Sensitivity Analysis. Frequency of recoverable format deviations during memory opera- tions is used. Higher values indicate greater reliance on fallback parsing due to inconsistent structured outputs. BackboneMethodAnswer ScoreFormat Error gpt-4o-mini SimpleMem0.2891.20% Nemori 0.78117.91% Qwen-2.5-3B SimpleMem 0.1024.82% Nemori0.44730.38% updates compared togpt-4o-mini. This “Silent Failure” implies that while the agent can converse fluently in the short term, its long-term memory becomes corrupted due to failed write operations. 2) Method Sensitivity: The impact of the back- bone varies by architecture complexity. Append- only systems are relatively robust, as they require minimal structured generation. In contrast, graph- based and episodic architectures are highly sen- sitive: extracting entities, constructing relations, and performing logical deduplication significantly increase format errors under weaker backbones, of- ten leading to structural instability or collapse in memory maintenance. 4.5 System Performance Evaluation While accuracy is critical, the practical viability of agentic memory is constrained by latency and cost. Unlike read-only RAG systems, agentic memory introduces maintenance operations such as memory extraction, updates, and consolidation. We measure the user-facing load as retrieval (T read ), covering search and traversal, and generation (T gen ), includ- ing context processing and token decoding, while discussing maintenance overhead qualitatively. In this section, we quantify user-facing latency (T read +T gen ) and offline scalability using Table 5, while highlighting maintenance as a hidden system- level bottleneck. Table 5: The "Agency Tax": Efficiency Profiling. We evaluate the trade-off between runtime user latency and offline construction cost. User Latency(T read + T gen ) dictates the interactive experience, while Construction Cost reflects the scalability and economic feasibility of the system. Note that Maintenance Cost is omitted as it is often handled asynchronously. Method User-Facing Latency (per turn)Construction Cost (Offline) Retrieval (T read )Generation (T gen )Total (s)Time (h)Tokens (k) Full ContextN/A1.7261.726N/AN/A LOCOMO (Maharana et al., 2024)0.4150.3680.7830.861,623 AMem (Xu et al., 2025c)0.0621.1191.18115.001,486 MemoryOS (Kang et al., 2025a)31.2471.12532.372 7.834,043 Nemori (Nan et al., 2025)0.2540.8751.1293.257,044 MAGMA (Jiang et al., 2026a)0.4970.9651.4627.282,725 SimpleMem (Liu et al., 2026)0.0091.0481.0573.451,308 MemSkill (Zhang et al., 2026a)0.0050.3010.306 0.601,796 4.5.1 Latency and Maintenance Trade-offs in MAG We analyze the end-to-end user-perceived latency (T read +T gen ) alongside the often-overlooked main- tenance overhead (T write ). Although Full Con- text eliminates retrieval cost, it incurs the high- est generation latency (T gen ≈ 1.73s), confirm- ing that large pre-fill computation increases time- to-first-token. Lightweight systems such as Sim- pleMem and LOCOMO achieve sub-second la- tency (< 1.1s) through efficient indexing, while MAGMA maintains a balanced profile (∼ 1.46s), adding modest overhead for graph traversal. In contrast, MemoryOS emerges as a clear bottleneck, with latency exceeding 32 seconds, suggesting that strict hierarchical paging (e.g., STM→LTM recur- sion) is impractical for interactive settings. Beyond user-facing latency, maintenance intro- duces a hidden scalability constraint. Append- only or lightweight systems incur lower up- date cost, whereas structured architectures (e.g., MAGMA, AMem) require graph restructuring and LLM-driven consolidation. Although often asyn- chronous, excessive maintenance overhead risks throughput collapse, where updates lag behind user interactions and memory becomes stale. Thus, structured memory demands robust asynchronous infrastructure to remain viable at scale. 4.5.2 Offline Scalability: Time and Token Economics Beyond online latency, we evaluate the offline cost of building the memory index. AMem requires approximately 15 hours for construction far slower than other baselines, suggesting super-linear update complexity (e.g., pairwise consolidation) that limits scalability on large datasets. Token consumption further exposes cost trade- offs. Nemori uses over 7.04M tokens during in- dex construction, nearly five times that of Sim- pleMem (1.3M). Although this yields strong ac- curacy, it reflects a substantial “intelligence tax,” where improved memory quality incurs signifi- cantly higher operational cost. In comparison, MAGMA achieves a more favorable Pareto bal- ance, delivering robust performance with moderate token usage (2.7M). 5 Conclusion and Future Directions Our analysis shows that agentic memory is limited not only by architecture, but also by evaluation validity, scalability, and robustness. 1. Rethinking Benchmark and Evaluation De- sign. Future benchmarks should be saturation- aware. As context windows expand, full-context prompting may solve many tasks, making it harder to isolate the value of external memory. Bench- marks should therefore stress task volume, tempo- ral depth, entity diversity, and long-range depen- dency, while using the Context Saturation Gap (∆) as a diagnostic signal. Evaluation should also move beyond lexical overlap. F1-style metrics often miss semantic cor- rectness, while LLM-as-a-judge requires prompt calibration and robustness checks. 2. Designing Scalable and Robust Agentic Mem- ory Systems. Agentic memory systems must balance accuracy, latency, cost, and reliability. Structured memory improves reasoning but intro- duces maintenance overhead, while lightweight ap- proaches improve efficiency but may lack abstrac- tion. Future systems should explicitly model write latency, maintenance throughput, and user-facing cost. They should also use backbone-aware mem- ory operations, constrained decoding, validation layers, or adaptive schemas to reduce silent corrup- tion. Limitations This survey has several limitations. First, agentic memory is evolving rapidly, so our taxonomy may miss concurrent or very recent systems. Second, our empirical analysis covers representative MAG architectures and selected benchmarks rather than all systems and settings; results may vary with im- plementations, prompts, model versions, and API behavior. These limitations call for broader, stan- dardized evaluations of agentic memory systems. References Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 techni- cal report. arXiv preprint arXiv:2303.08774. Alen Adamyan, Tomáš ˇ Cížek, Matej Straka, Klara Janouskova, and Martin Schmid. 2025. Sam2rl: Towards reinforcement learning memory control in segment anything model 2.arXiv preprint arXiv:2507.08548. Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901. Linyue Cai, Yuyang Cheng, Xiaoding Shao, Huim- ing Wang, Yong Zhao, Wei Zhang, and Kang Li. 2025. A scenario-driven cognitive approach to next-generation ai memory.arXiv preprint arXiv:2509.13235. Zouying Cao, Jiaji Deng, Li Yu, Weikang Zhou, Zhaoyang Liu, Bolin Ding, and Hai Zhao. 2025. Re- member me, refine me: A dynamic procedural mem- ory framework for experience-driven agent evolution. arXiv preprint arXiv:2512.10696. Guoxin Chen, Zile Qiao, Xuanzhong Chen, Donglei Yu, Haotian Xu, Wayne Xin Zhao, Ruihua Song, Wen- biao Yin, Huifeng Yin, Liwen Zhang, and 1 others. 2025. Iterresearch: Rethinking long-horizon agents via markovian state reconstruction. arXiv preprint arXiv:2511.07327. Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025. Mem0: Building production-ready ai agents with scalable long-term memory. arXiv preprint arXiv:2504.19413. Cody V Dong, Qihong Lu, Kenneth A Norman, and Se- bastian Michelmann. 2025. Towards large language models with human-like episodic memory. Trends in Cognitive Sciences. Xingbo Du, Loka Li, Duzhen Zhang, and Le Song. 2025a.MemR 3 : Memory retrieval via reflec- tive reasoning for LLM agents.arXiv preprint arXiv:2512.20237. Yiming Du, Baojun Wang, Yifan Xiang, Zhaowei Wang, Wenyu Huang, Boyang Xue, Bin Liang, Xing- shan Zeng, Fei Mi, Haoli Bai, and 1 others. 2025b. Memory-t1: Reinforcement learning for temporal reasoning in multi-session agents. arXiv preprint arXiv:2512.20092. Wenzhe Fan, Ning Yan, and Masood Mortazavi. 2025. Evomem: Improving multi-agent planning with dual- evolving memory. arXiv preprint arXiv:2511.01912. Jizhan Fang, Xinle Deng, Haoming Xu, Ziyan Jiang, Yuqi Tang, Ziwen Xu, Shumin Deng, Yunzhi Yao, Mengru Wang, Shuofei Qiao, and 1 oth- ers. 2025a. Lightmem: Lightweight and efficient memory-augmented generation.arXiv preprint arXiv:2510.18866. Runnan Fang, Yuan Liang, Xiaobin Wang, Jialong Wu, Shuofei Qiao, Pengjun Xie, Fei Huang, Hua- jun Chen, and Ningyu Zhang. 2025b. Memp: Ex- ploring agent procedural memory. arXiv preprint arXiv:2508.06433. Dongge Han, Camille Couturier, Daniel Madrigal Diaz, Xuchao Zhang, Victor Rühle, and Saravan Rajmohan. 2025. Legomem: Modular procedural memory for multi-agent llm systems for workflow automation. arXiv preprint arXiv:2510.04851. Chuanrui Hu, Xingze Gao, Zuyi Zhou, Dannong Xu, Yi Bai, Xintong Li, Hui Zhang, Tong Li, Chong Zhang, Lidong Bing, and 1 others. 2026a. Ever- memos: A self-organizing memory operating system for structured long-horizon reasoning. arXiv preprint arXiv:2601.02163. Yuyang Hu, Jiongnan Liu, Jiejun Tan, Yutao Zhu, and Zhicheng Dou. 2026b. Memory matters more: Event- centric memory as a logic map for agent searching and reasoning. arXiv preprint arXiv:2601.04726. Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang, Boyang Liu, Fangyi Zhu, Jiahang Lin, Honglin Guo, Shihan Dou, Zhiheng Xi, and 1 others. 2025. Memory in the age of ai agents. arXiv preprint arXiv:2512.13564. Wei-Chieh Huang, Weizhi Zhang, Yueqing Liang, Yuanchen Bei, Yankai Chen, Tao Feng, Xinyu Pan, Zhen Tan, Yu Wang, Tianxin Wei, and 1 others. 2026. Rethinking memory mechanisms of foun- dation agents in the second half. arXiv preprint arXiv:2602.06052. Yizhe Huang, Yang Liu, Ruiyu Zhao, Xiaolong Zhong, Xingming Yue, and Ling Jiang. 2025a. Memorb: A plug-and-play verbal-reinforcement memory layer for e-commerce customer service. arXiv preprint arXiv:2509.18713. Zhengjun Huang, Zhoujin Tian, Qintian Guo, Fangyuan Zhang, Yingli Zhou, Di Jiang, Zeying Xie, and Xi- aofang Zhou. 2025b. Licomemory: Lightweight and cognitive agentic memory for efficient long-term rea- soning. arXiv preprint arXiv:2511.01448. Yupeng Huo, Yaxi Lu, Zhong Zhang, Haotian Chen, and Yankai Lin. 2026. Atommem: Learnable dy- namic agentic memory with atomic memory opera- tion. arXiv preprint arXiv:2601.08323. Rodrigo Toro Icarte, Richard Valenzano, Toryn Q Klassen, Phillip Christoffersen, Amir-massoud Farah- mand, and Sheila A McIlraith. 2020. The act of re- membering: A study in partially observable reinforce- ment learning. arXiv preprint arXiv:2010.01753. Shian Jia, Ziyang Huang, Xinbo Wang, Haofei Zhang, and Mingli Song. 2025. Pisa: A pragmatic psych- inspired unified memory system for enhanced ai agency. arXiv preprint arXiv:2510.15966. Zixia Jia, Jiaqi Li, Yipeng Kang, Yuxuan Wang, Tong Wu, Quansen Wang, Xiaobo Wang, Shuyi Zhang, Junzhe Shen, Qing Li, and 1 others. 2026. The ai hippocampus: How far are we from human memory? arXiv preprint arXiv:2601.09113. Dongming Jiang, Yi Li, Guanpeng Li, and Bingzhe Li. 2026a. Magma: A multi-graph based agentic memory architecture for ai agents. arXiv preprint arXiv:2601.03236. Dongming Jiang, Yi Li, Guanpeng Li, Qiannan Li, and Bingzhe Li. 2026b. Hage: Harnessing agentic mem- ory via rl-driven weighted graph evolution. arXiv preprint arXiv:2605.09942. Hanqi Jiang, Junhao Chen, Yi Pan, Ling Chen, Weihang You, Yifan Zhou, Ruidong Zhang, Yohannes Abate, and Tianming Liu. 2026c. Synapse: Empowering llm agents with episodic-semantic memory via spreading activation. arXiv preprint arXiv:2601.02744. Jiazheng Kang, Mingming Ji, Zhe Zhao, and Ting Bai. 2025a. Memory os of ai agent. arXiv preprint arXiv:2506.06326. Jikun Kang,Wenqi Wu,Filippos Christianos, Alex James Chan, Fraser David Greenlee, George Thomas, Marvin Purtorab, and Andrew Toulis. 2025b. Lm2: Large memory models for long context reason- ing. In Workshop on Reasoning and Planning for Large Language Models. Minki Kang, Wei-Ning Chen, Dongge Han, Huseyin A Inan, Lukas Wutschitz, Yanzhi Chen, Robert Sim, and Saravan Rajmohan. 2025c. Acon: Optimiz- ing context compression for long-horizon llm agents. arXiv preprint arXiv:2510.00615. Sangyeop Kim, Yohan Lee, Sanghwa Kim, Hyunjong Kim, and Sungzoon Cho. 2025. Pre-storage reason- ing for episodic memory: Shifting inference burden to memory for personalized dialogue. arXiv preprint arXiv:2509.10852. Taeyoon Kwon, Dongwook Choi, Hyojun Kim, Sungh- wan Kim, Seungjun Moon, Beong-woo Kwak, Kuan- Hao Huang, and Jinyoung Yeo. 2025. Embodied agents meet personalization: Investigating challenges and solutions through the lens of memory utilization. arXiv preprint arXiv:2505.16348. Chris Latimer, Nicoló Boschi, Andrew Neeser, Chris Bartholomew, Gaurav Srivastava, Xuan Wang, and Naren Ramakrishnan. 2025. Hindsight is 20/20: Building agent memory that retains, recalls, and re- flects. arXiv preprint arXiv:2512.12818. Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. 2024. A human-inspired reading agent with gist memory of very long contexts. arXiv preprint arXiv:2402.09727. Haichang Li. 2025.Memory as a service (maas): Rethinking contextual memory as service-oriented modules for collaborative agents. arXiv preprint arXiv:2506.22815. Hao Li, Chenghao Yang, An Zhang, Yang Deng, Xi- ang Wang, and Tat-Seng Chua. 2025a. Hello again! llm-powered personalized agent for long-term dia- logue. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 5259– 5276. Kai Li, Xuanqing Yu, Ziyi Ni, Yi Zeng, Yao Xu, Zheqing Zhang, Xin Li, Jitao Sang, Xiaogang Duan, Xuelei Wang, and 1 others. 2026a. Timem: Temporal-hierarchical memory consolidation for long-horizon conversational agents. arXiv preprint arXiv:2601.02845. Yi Li, Lianjie Cao, Faraz Ahmed, Puneet Sharma, and Bingzhe Li. 2026b. Hippocampus: An efficient and scalable memory module for agentic ai. Preprint, arXiv:2602.13594. Zhiyu Li, Chenyang Xi, Chunyu Li, Ding Chen, Boyu Chen, Shichao Song, Simin Niu, Hanyu Wang, Jiawei Yang, Chen Tang, and 1 others. 2025b. Memos: A memory os for ai system. arXiv preprint arXiv:2507.03724. Zouying Cao Li Yu, Jiaji Deng. 2025. Agentscopereme: Memory management kit for agents. Shuai Ling, Lizi Liao, Dongmei Jiang, and Weili Guan. 2025. Memad: Structured memory of debates for en- hanced multi-agent reasoning. In Second Conference on Language Modeling. Genglin Liu, Shijie Geng, Sha Li, Hejie Cui, Sarah Zhang, Xin Liu, and Tianyi Liu. 2025a. Webcoach: Self-evolving web agents with cross-session memory guidance. arXiv preprint arXiv:2511.12997. Jiaqi Liu, Yaofeng Su, Peng Xia, Siwei Han, Zeyu Zheng, Cihang Xie, Mingyu Ding, and Huaxiu Yao. 2026. Simplemem: Efficient lifelong memory for llm agents. arXiv preprint arXiv:2601.02553. Jun Liu, Zhenglun Kong, Changdi Yang, Fan Yang, Tianqi Li, Peiyan Dong, Joannah Nanjekye, Hao Tang, Geng Yuan, Wei Niu, and 1 others. 2025b. Rcr- router: Efficient role-aware context routing for multi- agent llm systems with structured memory. arXiv preprint arXiv:2508.04903. Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paran- jape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language mod- els use long contexts. Transactions of the Association for Computational Linguistics, 12:157–173. Xinxin Liu, Weizhen Li, Weichen Sun, Xinlong Yang, Tianrui Qin, Xitong Gao, and Wangchunshu Zhou. 2025c. Compressed step information memory for end-to-end agent foundation models. Yaoyao Liu, Bernt Schiele, and Qianru Sun. 2021. Rmm: Reinforced memory management for class- incremental learning. Advances in neural informa- tion processing systems, 34:3478–3490. Jinghao Luo, Yuchen Tian, Chuxue Cao, Ziyang Luo, Hongzhan Lin, Kaixin Li, Chuyi Kong, Ruichao Yang, and Jing Ma. 2026. From storage to experi- ence: A survey on the evolution of llm agent memory mechanisms. Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. 2024.Evaluating very long-term conver- sational memory of llm agents.arXiv preprint arXiv:2402.17753. Wenyu Mao, Haosong Tan, Shuchang Liu, Haoyang Liu, Yifan Xu, Huaxiang Ji, and Xiang Wang. 2026. Bi-mem: Bidirectional construction of hierarchical memory for personalized llms via inductive-reflective agents. arXiv preprint arXiv:2601.06490. Ali Modarressi, Ayyoob Imani, Mohsen Fayyaz, and Hinrich Schütze. 2023. Ret-llm: Towards a general read-write memory for large language models. arXiv preprint arXiv:2305.14322. Hyungho Na, Yunkyeong Seo, and Il-chul Moon. 2024. Efficient episodic memory utilization of cooperative multi-agent reinforcement learning. arXiv preprint arXiv:2403.01112. Jiayan Nan, Wenquan Ma, Wenlong Wu, and Yize Chen. 2025. Nemori: Self-organizing agent mem- ory inspired by cognitive science. arXiv preprint arXiv:2508.03341. Leyi Ouyang. 2025.Can memory-augmented llm agents aid journalism in interpreting and fram- ing news for diverse audiences?arXiv preprint arXiv:2507.21055. Siru Ouyang, Jun Yan, I Hsu, Yanfei Chen, Ke Jiang, Zifeng Wang, Rujun Han, Long T Le, Samira Daruki, Xiangru Tang, and 1 others. 2025. Reasoningbank: Scaling agent self-evolving with reasoning memory. arXiv preprint arXiv:2509.25140. Charles Packer, Vivian Fang, Shishir_G Patil, Kevin Lin, Sarah Wooders, and Joseph_E Gonzalez. 2023. Memgpt: Towards llms as operating systems. Ofir Press, Noah A Smith, and Mike Lewis. 2021. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409. Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. 2025. Zep: a tempo- ral knowledge graph architecture for agent memory. arXiv preprint arXiv:2501.13956. Nazmus Sakib, Protoy Barai, Sifat Ishmam Parisa, and Anindya Iqbal. 2025. Memagent: A cache-inspired framework for augmenting conversational web agents with task-specific information. Miao Su, Yucan Guo, Zhongni Hou, Long Bai, Zixuan Li, Yufei Zhang, Guojun Yin, Wei Lin, Xiaolong Jin, Jiafeng Guo, and 1 others. 2026. Beyond dialogue time: Temporal semantic memory for personalized llm agents. arXiv preprint arXiv:2601.07468. Haoran Sun, Zekun Zhang, and Shaoning Zeng. 2025a. Preference-aware memory update for long-term llm agents. arXiv preprint arXiv:2510.09720. Weiwei Sun, Miao Lu, Zhan Ling, Kang Liu, Xuesong Yao, Yiming Yang, and Jiecao Chen. 2025b. Scaling long-horizon llm agent via context-folding. arXiv preprint arXiv:2510.11967. Haoran Tan, Zeyu Zhang, Chen Ma, Xu Chen, Quanyu Dai, and Zhenhua Dong. 2025a. Mem- bench: Towards more comprehensive evaluation on the memory of llm-based agents. arXiv preprint arXiv:2506.21605. Zhen Tan, Jun Yan, I-Hung Hsu, Rujun Han, Zifeng Wang, Long Le, Yiwen Song, Yanfei Chen, Hamid Palangi, George Lee, and 1 others. 2025b.In prospect and retrospect: Reflective memory man- agement for long-term personalized dialogue agents. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8416–8439. Xiangru Tang, Tianrui Qin, Tianhao Peng, Ziyang Zhou, Daniel Shao, Tingting Du, Xinming Wei, He Zhu, Ge Zhang, Jiaheng Liu, and 1 others. Agent kb: A hierarchical memory framework for cross-domain agentic problem solving. In ICML 2025 Workshop on Collaborative and Federated Agentic Workflows. Dehao Tao, Guoliang Ma, Yongfeng Huang, and Minghu Jiang. 2026. Membox: Weaving topic conti- nuity into long-range memory for llm agents. arXiv preprint arXiv:2601.03785. He Wang, Wenyilin Xiao, Songqiao Han, and Hailiang Huang. 2025a. Stockmem: An event-reflection mem- ory framework for stock forecasting. arXiv preprint arXiv:2512.02720. Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self- attention distillation for task-agnostic compression of pre-trained transformers. Advances in neural in- formation processing systems, 33:5776–5788. Yu Wang and Xi Chen. 2025. Mirix: Multi-agent mem- ory system for llm-based agents. arXiv preprint arXiv:2507.07957. Yu Wang, Ryuichi Takanobu, Zhiqi Liang, Yuzhen Mao, Yuanzhe Hu, Julian McAuley, and Xiaojian Wu. 2025b. Mem-α: Learning memory con- struction via reinforcement learning. arXiv preprint arXiv:2509.25911. Zixuan Wang, Bo Yu, Junzhe Zhao, Wenhao Sun, Sai Hou, Shuai Liang, Xing Hu, Yinhe Han, and Yim- ing Gan. 2025c. Karma: Augmenting embodied ai agents with long-and-short term memory systems. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 1–8. IEEE. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elic- its reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837. Chao Wu and Gang Li. 2025.Enhancing genera- tive agents in social simulations: Bridging memory, emotion, and governance for realistic social simu- lations. In Intelligent Systems Conference, pages 50–59. Springer. Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. 2024. Longmemeval: Benchmarking chat assistants on long-term interac- tive memory. arXiv preprint arXiv:2410.10813. Yaxiong Wu, Yongyue Zhang, Sheng Liang, and Yong Liu. 2025a. Sgmem: Sentence graph memory for long-term conversational agents.arXiv preprint arXiv:2509.21212. Zijun Wu, Yongchang Hao, and Lili Mou. 2025b. Tok- mem: Tokenized procedural memory for large lan- guage models. arXiv preprint arXiv:2510.00444. Derong Xu, Yi Wen, Pengyue Jia, Yingyi Zhang, Yichao Wang, Huifeng Guo, Ruiming Tang, Xiangyu Zhao, Enhong Chen, Tong Xu, and 1 others. 2025a. From single to multi-granularity: Toward long-term mem- ory association and selection of conversational agents. arXiv preprint arXiv:2505.19549. Haoran Xu, Jiacong Hu, Ke Zhang, Lei Yu, Yuxin Tang, Xinyuan Song, Yiqun Duan, Lynn Ai, and Bill Shi. 2025b.Sedm: Scalable self-evolving distributed memory for agents.arXiv preprint arXiv:2509.09498. Jiexi Xu. 2025. Memory management and contextual consistency for long-running low-code agents. arXiv preprint arXiv:2509.25250. Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Jun- tao Tan, and Yongfeng Zhang. 2025c.A-mem: Agentic memory for llm agents. arXiv preprint arXiv:2502.12110. Sikuan Yan, Xiufeng Yang, Zuchao Huang, Ercong Nie, Zifeng Ding, Zonggen Li, Xiaowen Ma, Kristian Ker- sting, Jeff Z Pan, Hinrich Schütze, and 1 others. 2025. Memory-r1: Enhancing large language model agents to manage and utilize memories via reinforcement learning. arXiv preprint arXiv:2508.19828. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024a. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115. Chang Yang, Chuang Zhou, Yilin Xiao, Su Dong, Luyao Zhuang, Yujing Zhang, Zhu Wang, Zijin Hong, Zheng Yuan, Zhishang Xiang, and 1 others. 2026a. Graph-based agent memory: Taxonomy, techniques, and applications. arXiv preprint arXiv:2602.05665. Chengyuan Yang, Zequn Sun, Wei Wei, and Wei Hu. 2026b. Beyond static summarization: Proactive memory extraction for llm agents. arXiv preprint arXiv:2601.04463. Hongkang Yang, Zehao Lin, Wenjin Wang, Hao Wu, Zhiyu Li, Bo Tang, Wenqiang Wei, Jinbo Wang, Zeyun Tang, Shichao Song, and 1 others. 2024b. memory 3 : Language modeling with explicit mem- ory. arXiv preprint arXiv:2407.01178. Xiaofang Yang, Lijun Li, Heng Zhou, Tong Zhu, Xiaoye Qu, Yuchen Fan, Qianshan Wei, Rui Ye, Li Kang, Yiran Qin, and 1 others. 2026c. Toward efficient agents: Memory, tool learning, and planning. arXiv preprint arXiv:2601.14192. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christo- pher D Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 conference on empiri- cal methods in natural language processing, pages 2369–2380. Yiqun Yao, Naitong Yu, Xiang Li, Xin Jiang, Xuezhi Fang, Wenjia Ma, Xuying Meng, Jing Li, Aixin Sun, and Yequan Wang. 2025. Egomem: Lifelong mem- ory agent for full-duplex omnimodal models. arXiv preprint arXiv:2509.11914. Rui Ye, Zhongwang Zhang, Kuan Li, Huifeng Yin, Zhengwei Tao, Yida Zhao, Liangcai Su, Liwen Zhang, Zile Qiao, Xinyu Wang, and 1 others. 2025. Agentfold: Long-horizon web agents with proactive context management.arXiv preprint arXiv:2510.24699. Hongli Yu, Tinghong Chen, Jiangtao Feng, Jiangjie Chen, Weinan Dai, Qiying Yu, Ya-Qin Zhang, Wei- Ying Ma, Jingjing Liu, Mingxuan Wang, and 1 others. 2025. Memagent: Reshaping long-context llm with multi-conv rl-based memory agent. arXiv preprint arXiv:2507.02259. Qianhao Yuan, Jie Lou, Zichao Li, Jiawei Chen, Yaojie Lu, Hongyu Lin, Le Sun, Debing Zhang, and Xianpei Han. 2025. Memsearcher: Training llms to reason, search and manage memory via end-to-end reinforce- ment learning. arXiv preprint arXiv:2511.02805. Dell Zhang, Yue Feng, Haiming Liu, Changzhi Sun, Jixiang Luo, Xiangyu Chen, and Xuelong Li. 2025a. Conversational agents: From rag to ltm. In Proceed- ings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Infor- mation Retrieval in the Asia Pacific Region, pages 447–452. Guibin Zhang, Muxin Fu, Guancheng Wan, Miao Yu, Kun Wang, and Shuicheng Yan. 2025b. G-memory: Tracing hierarchical memory for multi-agent systems. arXiv preprint arXiv:2506.07398. Guibin Zhang, Muxin Fu, and Shuicheng Yan. 2025c. Memgen: Weaving generative latent memory for self- evolving agents. arXiv preprint arXiv:2509.24704. Haozhen Zhang, Quanyu Long, Jianzhu Bao, Tao Feng, Weizhi Zhang, Haodong Yue, and Wenya Wang. 2026a. Memskill: Learning and evolving mem- ory skills for self-evolving agents. arXiv preprint arXiv:2602.02474. Kai Zhang, Xinyuan Zhang, Ejaz Ahmed, Hongda Jiang, Caleb Kumar, Kai Sun, Zhaojiang Lin, Sanat Sharma, Shereen Oraby, Aaron Colak, and 1 others. 2025d. Assomem: Scalable memory qa with multi-signal as- sociative retrieval. arXiv preprint arXiv:2510.10397. Ningning Zhang, Xingxing Yang, Zhizhong Tan, Weip- ing Deng, and Wenyong Wang. 2026b. Himem: Hi- erarchical long-term memory for llm long-horizon agents. arXiv preprint arXiv:2601.06377. Shengtao Zhang, Jiaqian Wang, Ruiwen Zhou, Jun- wei Liao, Yuchen Feng, Weinan Zhang, Ying Wen, Zhiyu Li, Feiyu Xiong, Yutao Qi, and 1 others. 2026c. Memrl: Self-evolving agents via runtime reinforce- ment learning on episodic memory. arXiv preprint arXiv:2601.03192. Xin Zhang, Kailai Yang, Hao Li, Chenyue Li, Qiyu Wei, and Sophia Ananiadou. 2026d. Implicit graph, explicit retrieval: Towards efficient and interpretable long-horizon memory for large language models. arXiv preprint arXiv:2601.03417. Yiran Zhang, Jincheng Hu, Mark Dras, and Usman Naseem. 2025e. Cogmem: A cognitive memory ar- chitecture for sustained multi-turn reasoning in large language models. arXiv preprint arXiv:2512.14118. Yuxiang Zhang, Jiangming Shu, Ye Ma, Xueyuan Lin, Shangxi Wu, and Jitao Sang. 2025f. Memory as ac- tion: Autonomous context curation for long-horizon agentic tasks. arXiv preprint arXiv:2510.12635. Xinkui Zhao, Qingyu Ma, Yifan Zhang, Hengxuan Lou, Guanjie Cheng, Shuiguang Deng, and Jianwei Yin. 2025. Ame: An efficient heterogeneous agen- tic memory engine for smartphones. arXiv preprint arXiv:2511.19192. Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. 2024. Memorybank: Enhancing large language models with long-term memory. In Pro- ceedings of the AAAI Conference on Artificial Intelli- gence, volume 38, pages 19724–19731. Huichi Zhou, Yihang Chen, Siyuan Guo, Xue Yan, Kin Hei Lee, Zihan Wang, Ka Yiu Lee, Guchun Zhang, Kun Shao, Linyi Yang, and 1 others. 2025a. Memento: Fine-tuning llm agents without fine-tuning llms. arXiv preprint arXiv:2508.16153. Sizhe Zhou and Jiawei Han. 2025. A simple yet strong baseline for long-term conversational memory of llm agents. arXiv preprint arXiv:2511.17208. Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Jinhua Zhao, Bryan Kian Hsiang Low, and Paul Pu Liang. 2025b. Mem1: Learning to synergize memory and reason- ing for efficient long-horizon agents. arXiv preprint arXiv:2506.15841. Yuanjie Zhu, Liangwei Yang, Ke Xu, Weizhi Zhang, Zihe Song, Jindong Wang, and Philip S Yu. 2025. Llm-memcluster: Empowering large language mod- els with dynamic memory for text clustering. arXiv preprint arXiv:2511.15424. A Taxonomy of Agentic Memory Figure 1 provides a comprehensive visual tax- onomy of recent advancements in Memory- Augmented Generation (MAG). Given the rapid proliferation of memory architectures for LLM agents, this tree diagram categorizes contemporary literature into four distinct structural paradigms: B Agentic Memory Background B.1 Memory Operations in Agentic Systems We characterize agentic memory as an external, non-parametric subsystem that interacts with an agent through two coupled processes: inference- time recall (reading memory to condition deci- sions) and memory update (writing, consolidating, and forgetting to maintain a useful long-term store). Letf θ denote a frozen foundation model (or policy model) with parametersθ, and letM t denote the external memory state at stept. Given an observa- tiono t (e.g., user input, tool output, sensor data) and agent states t (e.g., goals, plans, tool traces), the agent first produces a queryq t and recalls rele- vant memory: q t = Query(o t ,s t ),(3) r t = Read(M t ,q t ).(4) The retrieved contentr t is then integrated into the model input to produce an action or response: a t ∼ π θ (o t ,r t ,s t ) = f θ φ(o t ,s t ) ⊕ ψ(r t ) , (5) whereφ(·)formats current context,ψ(·)formats retrieved memory, and⊕denotes an integration operator (e.g., concatenation, schema-based slots, or cross-modal fusion). This abstraction makes explicit that agentic memory influences behavior through an external recall term rather than by up- dating θ. Inference-time retrieval as approximate utility optimization. External memory retrieval typi- cally selects itemsm i N i=1 fromM t using a scor- ing function (dense similarity, sparse matching, or reranking). A common instantiation is top-krecall: r t = TopK m i N i=1 ; score(q t ,m i ), k .(6) However, in agentic settings, the ideal “relevance” is not purely semantic but decision-conditional. One can express an idealized retrieval objective as selecting memory that maximizes downstream utility: r ⋆ t = arg max r⊆M t E U (a t | o t ,r,s t ) ,(7) whereU (·)denotes agent utility (e.g., task success, efficiency, robustness). Practical systems approxi- mate(7)using similarity search, learned rerankers, multi-hop retrieval, planner-guided recall, or re- trieval policies trained to better alignscore(·)with utility. Memory update as explicit memory actions. After producing an actiona t (and possibly observ- ing its outcome), the agent updates external mem- ory through a write function: M t+1 = Write(M t ,o t ,a t ,s t ).(8) It is often useful to make updates explicit as mem- ory actions. Letu t ∈ Udenote a memory ac- tion such as STORE, UPDATE, SUMMARIZE, LINK, EVICT, or DELETE. Then: u t = g(o t ,a t ,s t ), M t+1 = T (M t ,u t ), (9) whereg(·)may be rule-based, model-driven, or learned, andTapplies the chosen action to the memory store. This view connects naturally to RL- guided memory, whereg(·)can be optimized as a policy over memory actions. C Related Work Several recent surveys have examined memory mechanisms for Agentic AI systems, each from a distinct vantage point. The AI Hippocampus (Jia et al., 2026) presents a broad synthesis organized around a brain-inspired trichotomy of implicit memory, explicit memory, and agentic memory, fur- ther extending the analysis to multimodal settings involving vision, audio, and embodied interaction. Memory in the Age of AI Agents (Hu et al., 2025) proposes a “forms–functions–dynamics” frame- work that categorizes agent memory along three or- thogonal axes: architectural form, functional role, and lifecycle dynamics, providing a comprehen- sive conceptual vocabulary for the rapidly frag- menting landscape of agent memory research. To- ward Efficient Agents (Yang et al., 2026c) shifts the focus from architectural expressiveness to de- ployment cost, surveying efficiency-oriented tech- niques across three core agent components: mem- ory, tool learning, and planning. In addition, this Memory-Augmented Generation (MAG) Lightweight Semantic Memory RL-Optimized Semantic Compression MemAgent (Yu et al., 2025), MemSearcher (Yuan et al., 2025), Mem-α(Wang et al., 2025b), Memory-R1 (Yan et al., 2025) Context Window Management AgentFold (Ye et al., 2025), Context-Folding Agent (Sun et al., 2025b), LLM- MemCluster (Zhu et al., 2025), MemAgent-Cache (Sakib et al., 2025), Light- mem (Fang et al., 2025a), Lm2 (Kang et al., 2025b), RCR-Router (Liu et al., 2025b) Heuristic / Prompt-Optimized SimpleMem (Liu et al., 2026), ACON (Kang et al., 2025c), CISM (Liu et al., 2025c), AME (Zhao et al., 2025), CogMem (Zhang et al., 2025e), Pisa (Jia et al., 2025), Memento (Zhou et al., 2025a) Token-Level Semantic Memory TokMem (Wu et al., 2025b), MemGen (Zhang et al., 2025c),Memory 3 (Yang et al., 2024b) Entity-Centric & Personalized Memory Entity-Centric Memory RET-LLM (Modarressi et al., 2023), A-MEM (Xu et al., 2025c), Memory-R1 (Yan et al., 2025), Mem0 (Chhikara et al., 2025), RMM (Liu et al., 2021), Web- Coach (Liu et al., 2025a), MEM1 (Zhou et al., 2025b) Personalized Memory PAMU (Sun et al., 2025a), EgoMem (Yao et al., 2025), MemOrb (Huang et al., 2025a), WebCoach (Liu et al., 2025a), MemoryBank (Zhong et al., 2024), Hello Again (Li et al., 2025a), O-Mem (Wang et al., 2025b), Embodied Agents (Kwon et al., 2025), Bi-Mem (Mao et al., 2026), Beyond Dialogue Time (Su et al., 2026) Episodic & Reflective Memory Episodic Buffer with Learned Control MemR3 (Du et al., 2025a), MemAct (Zhang et al., 2025f), Act of Remember- ing (Icarte et al., 2020) Episodic Reflection & Consolidation In Prospect & Retrospect (Tan et al., 2025b), TiMem (Li et al., 2026a), Memp (Fang et al., 2025b), Pre-Storage Reasoning (Kim et al., 2025), Nemori (Nan et al., 2025), LEGOMem (Han et al., 2025), Reasoning- Bank (Ouyang et al., 2025), ReadAgent (Lee et al., 2024), ProMem (Yang et al., 2026b), Human-like Episodic Memory (Dong et al., 2025), MemSkill (Zhang et al., 2026a) Episodic Recall for Exploration EMU (Na et al., 2024), SAM2RL (Adamyan et al., 2025) Episodic Utility Learning MemRL (Zhang et al., 2026c), Memory-T1 (Du et al., 2025b), Memento (Zhou et al., 2025a), Remember Me Refine Me (Cao et al., 2025) Structured & Hierarchical Memory Graph-Structured Memory MAGMA (Jiang et al., 2026a), Zep (Rasmussen et al., 2025), SGMem (Wu et al., 2025a), SYNAPSE (Jiang et al., 2026c), LatentGraphMem (Zhang et al., 2026d), LiCoMemory (Huang et al., 2025b), Hindsight (Latimer et al., 2025), Sim- ple Baseline (Zhou and Han, 2025), Multi-Granularity (Xu et al., 2025a), Stock- Mem (Wang et al., 2025a), G-Memory (Zhang et al., 2025b), Membox (Tao et al., 2026), Memory Matters More (Hu et al., 2026b), AssoMem (Zhang et al., 2025d), HAGE (Jiang et al., 2026b) OS-Inspired & Hierarchical Memory MemGPT (Packer et al., 2023), MemoryOS (Kang et al., 2025a), Ever- MemOS (Hu et al., 2026a), HiMem (Zhang et al., 2026b), MeMAD (Ling et al., 2025), Mirix (Wang and Chen, 2025), Scenario-Driven (Cai et al., 2025), ReMe (Li Yu, 2025), LightMem (Fang et al., 2025a), RAG-to-LTM (Zhang et al., 2025a), MemOS (Li et al., 2025b), EvoMem (Fan et al., 2025), AGENT KB (Tang et al.), KARMA (Wang et al., 2025c), MaaS (Li, 2025), CogMem (Zhang et al., 2025e), PISA (Jia et al., 2025), Memory 3 (Yang et al., 2024b),MADES (Ouyang, 2025), TiMem (Li et al., 2026a), (Xu, 2025) Policy-Optimized Memory Management MEM1 (Zhou et al., 2025b), Mem-α(Wang et al., 2025b), AtomMem (Huo et al., 2026), SEDM (Xu et al., 2025b), Memory-R1 a(Yan et al., 2025), Memory-T1 (Du et al., 2025b), IterResearch (Chen et al., 2025), RCR-Router (Liu et al., 2025b), LM2 (Kang et al., 2025b), MemSkill (Zhang et al., 2026a), HAGE (Jiang et al., 2026b) Figure 1: Taxonomy of Memory-Augmented Generation (MAG) systems. survey discusses compression, context manage- ment, and reinforcement-learning-based reward de- sign as shared principles to reduce latency, token consumption, and interaction steps. More recently, Rethinking Memory Mechanisms (Huang et al., 2026) assembles a large-scale survey of over 200 papers, organizing memory along three dimensions: substrate, cognitive mechanism, and subject, while reviewing learning policies over memory opera- tions and cataloguing existing evaluation bench- marks. From Storage to Experience (Luo et al., 2026) offers an evolutionary perspective, formal- izing memory development into three progressive stages: storage, reflection, and experience. It also identifies long-range consistency, dynamic environ- ments, and continual learning as the core drivers of this evolution. Graph-based Agent Memory (Yang et al., 2026a) narrows the scope to graph-based memory paradigms of knowledge graphs, temporal graphs, hypergraphs, and hierarchical trees. In addi- tion, it systematically analyzes extraction, storage, retrieval, and evolution along the memory lifecycle. D Prompt Library This section details the prompt templates used for all experimental evaluations. To ensure repro- ducibility, we provide the specific instructions for memory construction, query processing, and the varying sensitivities of our evaluation protocols. To provide a structured comparison, we classify the prompt designs into three operational stages: Memory Construction (Build), Query Processing (Query), and Response Generation (Answer). Ta- ble 6 summarizes the design patterns across the evaluated systems. D.1 Memory Construction and Retrieval Different memory architectures require different construction strategies. This section outlines the prompts used by MAG systems to consolidate raw interaction history into long-term storage and refine user queries. Build Prompts (Memory Indexing) Used by MAG systems to consolidate raw interaction his- tory into long-term storage. The evaluated systems Table 6: Taxonomy of System Operation Prompts across Memory Architectures. SystemBuild Strategy (Memory Construction)Query StrategyAnswer Strategy (Synthesis) MemoryOSProfile-based (User profiling, Knowledge extraction)N/A (Direct Semantic Search)Role-playing & Profile-enriched AMemFlat/Turn-based (Content analysis)LLM Keyword ExtractionRetrieved Memory Context NemoriEpisodic (Boundary detection, Episode generation)N/A (Direct Semantic Search)Episode-based Retrieval MAGMAGraph-based (Event extraction, Multi-hop reasoning)Multi-hop Entity ParsingGraph Traversal Synthesis SimpleMemMinimalist/Turn-levelKeyword GenerationContext-based utilize distinct structural representations: •Profile Based (MemoryOS): Instructs the LLM to extract observable user traits and merge them into an evolving profile. •Episodic (Nemori): Segments continuous di- alogue into discrete episodes using boundary detection. •Graph Based (MAGMA): Translates inter- actions into relational structures (e.g., event extraction, triplet arrays). [PLACEHOLDER: Insert Build Prompt for memory consolidation, e.g., "Summa- rize the following interaction into atomic facts..."] Query Refinement Prompts While many sys- tems (like Nemori and MemoryOS) bypass explicit refinement in favor of direct semantic search al- gorithms, systems like AMem and SimpleMem use LLMs to transform user queries into optimized search vectors or keywords. [PLACEHOLDER: Insert Query Prompt, e.g., "Given the conversation history, rewrite the user query for better re- trieval..."] D.2 Response Generation Answer Generator Prompts The standard tem- plates used by all baselines (RAG, MAG, and Full- Context) to produce final responses based on re- trieved or provided context. The prompt designs vary based on the context strategy (e.g., profile- enriched role-playing for MemoryOS, episodic re- trieval for Nemori) and specific constraints (e.g., temporal awareness or multi-hop reasoning for AMem and MAGMA). [PLACEHOLDER: Insert Answer Gen- erator Prompt, e.g., "You are an assis- tant with access to the following mem- ory shards. Answer the question based on..."] D.3 LLM-as-a-Judge Evaluation Protocols We utilizegpt-4o-minias our backbone judge. To comprehensively evaluate architecture perfor- mance across the diverse grading criteria men- tioned in Section 4.3.2, we structure our evalua- tion into two categories: literature-derived baseline prompts and sensitivity rubrics. D.3.1 Literature-Derived Baselines These prompts represent different community stan- dards for "correctness" and are sourced directly from existing benchmarks. Prompt 1: MAGMA (Semantic Correctness & Context) Derived from the MAGMA frame- work (Jiang et al., 2026a), this multi-level scoring protocol prioritizes information integration and rea- soning. It emphasizes interpersonal knowledge retrieval and semantic equivalence, with specific guidelines for temporal and factual preservation. Score the answer on a scale from 0.0 to 1.0 based on semantic correctness. Scoring Scale: - 1.0: Perfect match — contains all key information, semantically equivalent - 0.8: Mostly correct — captures main point but may have minor differences - 0.6: Partially correct — has some correct info but incomplete - 0.4: Somewhat related — touches on topic but misses significant info - 0.2: Barely related — answer is mostly incorrect - 0.0: Completely wrong — answer is unrelated or contradicts gold answer Instruction: Focus on user-interpersonal knowledge and temporal generosity. Focus on semantic equivalence, not exact wording. Assign partial credit for partially correct answers. Input: Question: question Gold answer: gold_answer Generated answer: generated_answer Output (JSON): "score": 1.0, "reasoning": "..." Prompt 2: Nemori (Generous Semantic Match- ing)Adapted from the Nemori paper (Nan et al., 2025), this is a lenient, semantics-oriented eval- uation scheme. It emphasizes entity recall and judges whether the generated answer captures the same underlying concept as the ground truth using a binary (CORRECT/WRONG) classification, tolerating paraphrasing and verbosity. Your task is to label an answer as CORRECT or WRONG. You will be given: (1) a question (2) a gold (ground truth) answer (3) a generated answer Evaluation Guidelines: - Be generous in grading. - If the generated answer conveys the same meaning or topic as the gold answer, mark it as CORRECT. - Ignore differences in wording, phrasing, or length. - Accept paraphrases and semantically equivalent answers. - For time-related questions, accept different formats (e.g., “May 7” vs “7 May”). Input: Question: question Gold Answer: gold_answer Generated Answer: generated_answer First, provide a one-sentence reasoning, then output the result. Output (JSON): "label": "CORRECT" or "label": "WRONG" Prompt 3: SimpleMem (Relevance & Accuracy) Adapted from the SimpleMem baseline (Liu et al., 2026), this prompt focuses on retrieval precision and core fact preservation. It explicitly balances relevance, factual grounding, and tolerance to rep- resentational variation via a Robustness Protocol. You are an expert Relevance & Accuracy Evaluator. Your task is to determine whether the Predicted Answer successfully retrieves the necessary information to answer the Question, based on the Reference Answer. Input: Question: question Reference Answer: reference Predicted Answer: prediction Evaluation Criteria: 1. Responsiveness to Query The predicted answer must directly address the specific question and remain topically aligned with the user’s intent. 2. Core Fact Preservation The prediction must capture the key signal or core entity from the reference (e.g., who, what, or outcome). 3. Informational Utility The answer must provide meaningful value. Even if concise, it should convey the essential information required by the question. 4. Robustness Protocol (Acceptable Variances) You must treat the following variations as valid matches: - Temporal & numerical tolerance (e.g., ±1–2 days, rounded numbers) - Granularity differences (e.g., “Afternoon” vs. “14:05”, “Late October” vs. “Oct 25”) - Information subsetting (partial but sufficient answers) - Synonymy and format variation Grading Logic: - Score 1.0 (Pass): Contains relevant core information OR satisfies robustness conditions above. - Score 0.0 (Fail): Missing core information, irrelevant, or fails to answer the question. Output Format (JSON only): "score": 1.0, "reasoning": "Brief explanation focusing on relevance and core match." E Baseline Configurations This section details the hyper-parameter settings and model versions for the evaluated architec- tures. To ensure fair and objective comparisons, we strictly follow the default configuration settings provided in their respective open-source reposito- ries, with the following standardized modifications applied across all baseline systems: •Embedding Model: All dense retrieval op- erations are uniformly configured to use all-MiniLM-L6-v2(Wang et al., 2020), re- placing any system-specific default embed- ding models to ensure a controlled baseline for semantic matching. •LLM Temperature: The generation tempera- ture is fixed at0.3across all backbone LLMs to maintain a consistent balance between de- terminism and reasoning capability. • Retrieval Top-k: For final answer synthesis that relies on retrieving raw conversation his- tory or memory chunks, we uniformly set the retrieval scope to top-k = 10. • Max Tokens: The maximum token limits for generation and context windows are main- tained at their repository-specific defaults to respect the intended design of each architec- ture. A summary of these unified hyper-parameters alongside the specific operational parameters for each evaluated system is provided in Table 7. F Case Studies: Why Lexical Metrics Fail To further investigate the ranking discrepancies ob- served in Section 4.3.2, we conduct a qualitative analysis of representative cases where lexical met- rics (e.g., F1) disagree with semantic judgments. Rather than presenting isolated examples, we or- ganize these cases into a set of recurring failure mechanisms that reflect inherent limitations of to- ken level evaluation. We identify four common patterns: •Surface Variation: Correct answers ex- pressed with additional context or alternative phrasing are penalized due to reduced lexical overlap. •Semantic Equivalence Gap: Equivalent meanings conveyed through different formats or synonyms result in zero or near-zero scores. •Polarity Flip: Minor lexical changes (e.g., negation) invert the semantic meaning while preserving high token overlap. •Entity Drift: Incorrect entities or values are substituted within otherwise similar sentence structures, leading to inflated lexical similarity despite factual errors. Table 9 presents representative examples illus- trating these failure modes. These cases demon- strate that lexical metrics are not merely noisy, but systematically misaligned with the abstraction, nor- malization, and reasoning behaviors exhibited by modern LLM-based systems. These observations complement the quantitative findings in Section 4.3.2 and suggest a fundamental mismatch: lexical metrics operate on surface form alignment, whereas agentic systems increasingly rely on abstraction, normalization, and composi- tional reasoning. As a result, improvements in rea- soning quality may not be reflected and can even be penalized by traditional token-based evaluation. Table 7: Key hyper-parameter configurations for the evaluated memory architectures. To ensure fair comparison, embedding models, LLM generation temperatures, and final answer retrieval scopes are strictly standardized across all baselines, while structural capacity parameters (e.g., max tokens) follow repository defaults. MethodEmbedding ModelLLM Temp.Final Answer Top-kSystem-Specific Defaults (Max Tokens & Structure) Full ContextN/A0.3N/AMax Tokens: 128k (gpt-4o-mini) LOCOMOMiniLM-L6-v20.310Max Tokens: Default; Buffer Size: Default AMem MiniLM-L6-v20.310Max Tokens: Default; Keyword Extractor Temp: Default MemoryOSMiniLM-L6-v20.310Max Tokens: Default; Update Frequency: Default NemoriMiniLM-L6-v20.310Max Tokens: Default; Boundary Temp: 0.1 MAGMAMiniLM-L6-v20.310Max Tokens: Default; Consolidation Threshold: Default SimpleMem MiniLM-L6-v20.310Max Tokens: Default; Synthesis Strategy: Default Table 8: Overview of memory systems and experimental configurations. We usegpt-4o-minias the primary controller for all methods in the main benchmark to normalize reasoning costs. MethodMemory StructureUpdate PolicyRetrieval Scope A-MEMLinked Node Graph (Atomic Units + Tags) Evolutionary:LLM-based node rewriting & dynamic linking Dense Embedding Similarity (Top-k) MemoryOS Hierarchical Tiers (STM→LTM)Rule-based:Frequency/Recency- based promotion Cascading Hierarchy Search MAGMAMulti-relational Graph + Vector Index Asynchronous: Dual-stream consoli- dation (Long-term) Intent-guided Subgraph Traver- sal Nemori Dual Memory (Episodic Tree + Semantic Graph) Gradient-inspired: Contextual mem- ory modification Hybrid (Top-kEpisodes + Se- mantic Facts) SimpleMemHybrid Index (Dense/Sparse) of Compressed Units Synchronous: On-the-fly synthesis & deduplication Planner-guidedMulti-view Search Table 9: Mechanism Oriented Failure Cases of Lexical Metrics. Lexical scores (F1) are contrasted with semantic judgments to highlight systematic mismatches. Query & Gold TruthModel AnswerFailure TypeF1JudgeAnalysis Q: What is the duration of the event? Gold: 18 days The total duration was 18 days. Surface Variation0.501.00Additional phrasing lowers lex- ical precision despite identical meaning. Q: What time does the event start? Gold: 14:00 2 PMSemantic Equiva- lence Gap 0.001.00Equivalent time representations yield zero token overlap. Q: Describe the price level. Gold: cheap inexpensiveSemantic Equiva- lence Gap 0.001.00Synonym substitution is not cap- tured by lexical matching. Q: Is the software compatible with Mac? Gold: compatible with Mac not compatible with Mac Polarity Flip0.8570.00Negation reverses meaning while preserving token overlap. Q: Who completed the project? Gold: John completed the project Sarah completed the project Entity Drift0.750.00Incorrect entity maintains struc- ture but changes semantics. Q: How many items were included? Gold: three items five itemsEntity Drift0.500.00Numerical substitution is par- tially rewarded due to shared to- kens.

Instant research discovery