WorksComputer Science
Learning to Remember: Advances in Memory Mechanisms for Large Language Models
Abstract
Memory has become a central mechanism for extending large language models (LLMs) beyond fixed context windows, stateless interaction, and static parametric knowledge. This survey organizes the literature around three complementary dimensions: the life cycle of memory, the storage substrate, and the system role that memory plays in LLM-based agents, while using cognitive-science-inspired memory concepts as a lightweight conceptual motivation rather than as a separate technical contribution or implementation family. On this basis, we distinguish three mechanism families: nonparametric memory, where information is stored outside model parameters in text, graph, multimodal, or database-like structures; parametric memory, where information is stored or accessed through activations, key-value caches, learned modules, adapter-based parameterizations, or model weights; and memory for agent evolution, where stored experiences support reinforcement learning, continual learning, procedural reuse, context engineering, and agent harness design. Evaluation is treated as a separate cross-cutting component rather than as a memory mechanism. Unlike surveys that focus primarily on retrieval-augmented generation or long-context modeling, this paper emphasizes the full memory life cycle: construction, maintenance, retrieval, consolidation, forgetting, governance, and evaluation. By aligning conceptual motivations, computational memory mechanisms, benchmarks, and deployment concerns, the survey provides a structured roadmap for designing memory-augmented LLM agents that are persistent, adaptive, auditable, and safe to update over time.
keywords: Large language model; Memory mechanism; Agent memory; Retrieval-augmented generation; Long-term memory
Introduction
The rapid advancement of large language models (LLMs) has reshaped natural language processing and artificial intelligence. Scaling has improved reasoning, comprehension, and generation [14, 145], yet conventional LLMs remain constrained by fixed context windows, static parameters, and weak mechanisms for preserving state across interactions. These limitations make it difficult for models to maintain continuity, accumulate experience, adapt to changing user needs, or reason over long-horizon tasks [15]. Memory mechanisms have therefore become a key research direction for transforming LLMs from stateless generators into adaptive agents.
Introducing memory into LLMs enables models to store, retrieve, and revise information across time. In dialogue and personalization, memory supports persistent user profiles, long-term preferences, and temporal consistency, as illustrated by MemoryBank [222] and COMEDY [25]; in systems with explicit lifecycle operations, such as Mem0 [31], memory can also support correction or updating of outdated facts. In knowledge-intensive settings, memory helps agents reuse prior evidence, maintain structured knowledge, and reduce repeated retrieval. In long-horizon agent systems, memory further supports planning, reflection, skill reuse, and behavioral improvement across repeated tasks.
This survey organizes LLM memory research into a lightweight conceptual-motivation layer, three computational mechanism families, and one cross-cutting evaluation component. The conceptual layer uses cognitive-science-inspired distinctions such as episodic, semantic, working, procedural, consolidated, and context-dependent memory only to clarify the vocabulary for comparing AI memory designs; it is not treated as a standalone mechanism family or the central contribution of this survey. Among computational mechanisms, nonparametric memory stores information outside model parameters and includes text memory, graph memory, database-backed memory, and multimodal memory. Parametric memory stores or accesses knowledge through internal states, KV caches, trainable memory modules, adapter-based parameterizations, or direct parameter updates. Memory for agent evolution treats memory as a reusable experience substrate for reinforcement learning, continual learning, procedural reuse, context engineering, and agent harness design. Separately, evaluation studies how these memory mechanisms are measured in plain-text, multimodal, and simulation-based environments. This distinction avoids treating cognitive analogies or benchmarks as additional storage substrates while still making both conceptual grounding and evaluation central to the survey.
The central contribution of this survey is a conceptual-motivation–lifecycle–substrate–role–evaluation alignment framework for comparing LLM memory research. Instead of grouping systems only by implementation type, this framework aligns five questions: which conceptual memory distinctions motivate the design without themselves constituting a memory mechanism, what lifecycle operations a memory mechanism supports, which substrate realizes the memory, what role the memory plays in an LLM-based system, and how the claimed benefit is evaluated. On this basis, the contributions of this survey are threefold. (1) We provide a consolidated map of LLM memory systems that uses cognitive-science-inspired concepts only as a compact vocabulary for motivating three computational memory-mechanism families and the evaluation resources used to test them. (2) We analyze representative methods through a shared set of lifecycle questions: what is written, where it is stored, how it is retrieved, how it is updated or forgotten, how persistent it is, and how it is governed. (3) We synthesize benchmark trends and future challenges, highlighting where existing systems still struggle with reliable writing, temporal updates, privacy, deletion, hallucination control, and long-horizon utility. This alignment-oriented framing makes the relationships among conceptual motivation, memory substrates, system roles, evaluation protocols, and deployment risks explicit, thereby providing a common basis for comparing otherwise heterogeneous memory-augmented LLM systems.
Scope and Literature Selection
This survey focuses on methods in which memory is an explicit design object rather than a generic byproduct of long prompts or model scale. The literature was selected by combining keyword-driven search with backward and forward citation tracing. Search terms included combinations of LLM memory, agent memory, long-term memory, episodic memory, semantic memory, procedural memory, graph memory, multimodal memory, KV cache, model editing, memory benchmark, and self-evolving agents. The search covered peer-reviewed venues, preprints, benchmark papers, benchmark repositories, system documentation, and project pages when they introduced a concrete memory architecture or operational interface.
We include systems that specify how memory is written, stored, retrieved, updated, compressed, forgotten, governed, or evaluated. The reviewed literature covers external memory stores, graph-based memory, multimodal memory, KV-cache and activation memory, weight- or module-level memory, self-evolving agents, cognitive-science memory analogies, and benchmarks that test memory persistence or long-horizon behavior. A method is included when memory is central to the mechanism or evaluation claim, not merely an incidental implementation detail. We exclude methods that only perform one-off retrieval, prompt compression, or long-context inference unless they directly address memory persistence, lifecycle management, update behavior, or memory-guided future action. This scope makes the paper a mechanism-centered survey rather than a general survey of RAG, long-context modeling, or agent frameworks.
Memory, RAG, and Long Context
Memory overlaps with retrieval-augmented generation (RAG) and long-context modeling, but it is not identical to either. RAG is primarily retrieval-centered: the original RAG formulation retrieves Wikipedia passages from a dense vector index with a neural retriever, while later RAG-style systems may use sparse, dense, or hybrid retrieval [84, 75]. Long-context modeling is capacity-centered: the system attempts to process more tokens within a single inference context. Memory is lifecycle-centered: it requires decisions about what to store, how to represent it, when to update it, how to resolve conflicts, how to delete it, and how to evaluate whether it improves future behavior. This distinction is important because a system may have a long context window or a retriever without having reliable long-term memory.
The remainder of this paper is organized as follows. § 2 introduces the organizing framework and memory life cycle. § 3 briefly explains the cognitive-science-inspired vocabulary used only as conceptual motivation. § 4 reviews nonparametric memory. § 5 discusses parametric memory. § 6 surveys memory for agent evolution. § 7 reviews benchmarks and evaluation environments. § 8 discusses open challenges, followed by the conclusion.
Overview
Research on memory in LLMs has expanded from early long-context retrieval and dialogue summarization to an ecosystem of memory stores, memory operating systems, graph memories, multimodal memories, parametric memories, and self-evolving agents. To organize this literature, we use cognitive-science-inspired memory concepts only as conceptual motivation, and compare computational methods along three complementary dimensions: the life cycle of memory, the form or substrate of memory, and the system role played by memory.
Based on these dimensions, Figure 2 presents the taxonomy used throughout this survey. The taxonomy now separates a compact conceptual motivation layer from the computational mechanism families. Cognitive-science-inspired concepts motivate how memory can be decomposed and evaluated, but the computational analysis centers on three families: nonparametric memory, parametric memory, and memory for agent evolution. Evaluation is shown alongside these families as a cross-cutting component because benchmarks determine whether a proposed memory mechanism actually improves persistence, update behavior, grounding, and long-horizon utility. Table 1 further summarizes the substrate, system role, and reviewed subcategories of each survey component, while later classification tables provide representative systems and citations.
| Survey component | Where memory resides | Main role in the system | Reviewed subcategories |
| Conceptual motivation | Cognitive-science theories and human-memory analogies rather than a computational storage substrate | Provides the conceptual vocabulary for separating memory types, lifecycle operations, and update signals | Episodic, semantic, working, and procedural memory; consolidation and forgetting; context binding; prediction-error updating |
| Nonparametric memory | External stores such as text repositories, vector databases, knowledge graphs, and multimodal archives | Provides an explicit, inspectable, and updateable memory layer outside model weights | Text memory; graph memory; multimodal memory for understanding and generation |
| Parametric memory | Internal model components such as activations, KV caches, adapter-based parameterizations, trainable modules, and model weights | Stores or injects memory within the model’s internal computation or parameters | Activation/KV construction; KV management; weight editing; learned memory modules |
| Memory for agent evolution | Interaction traces, rewards, failures, workflows, code traces, and context states | Supports behavioral improvement through experience accumulation, reflection, and skill reuse | RL and continual learning; procedural and skill memory; context and harness engineering |
| Evaluation | Benchmarks built from textual histories, videos, images, app worlds, and simulated environments | Measures memory reliability, persistence, update ability, and long-horizon utility | Plain-text benchmarks; multimodal benchmarks; simulation environments |
Table 1. Survey-level organization of LLM memory research. The table summarizes one conceptual-motivation layer, three computational memory-mechanism families, and one cross-cutting evaluation component, including the main substrate or target, system-level role, and reviewed subcategories.

Figure 2. Revised taxonomy of LLM memory mechanisms. Cognitive-science-inspired concepts are used only as conceptual motivation and terminology, while the main body of the survey is organized around memory lifecycle operations, three computational memory families, and evaluation.
Memory Life Cycle
Although implementations differ, most memory systems can be analyzed through five recurring operations. Figure 1 summarizes these operations as a lifecycle from memory construction to governed reuse in the LLM context.
Figure 1. LLM memory life cycle. Memory is constructed from source documents, conversations, perceptual streams, or model states; maintained through deduplication and conflict resolution; retrieved into the active context; consolidated or forgotten over time; and governed through access control, privacy filtering, and auditability.
Construction transforms raw interactions, documents, visual streams, or model states into memory units. Text systems summarize conversations into facts or episodes [222, 31]; graph systems extract entities,
relations, and events ([71, 126]); multimodal systems segment long videos or screen streams into captioned events ([96, 154]).
Management maintains memory quality through compression, deduplication, conflict resolution, forgetting, temporal decay, permission control, versioning, and cross-session synchronization. Mem0 ([31]) and MemOS ([92]) make memory maintenance explicit through update operations and lifecycle governance.
Retrieval selects relevant memory and reintroduces it into the active reasoning context. Retrieval may rely on sparse retrieval such as BM25 ([129]), dense neural retrieval such as DPR ([75]) or the original RAG ([84]) retriever, later hybrid retrieval, graph traversal, temporal filters, symbolic queries, or LLM-guided search ([147]).
Consolidation and forgetting decide which memories should become durable, be merged into higher-level abstractions, or be discarded. This operation is increasingly important in long-running agents, where unconstrained accumulation causes noise, privacy risk, and retrieval degradation.
Evaluation and governance measure whether memory improves downstream behavior and whether the system remains safe, private, and auditable. This dimension connects technical design to benchmark design and deployment requirements.
Foundations, Forms, and Roles of Memory
The second dimension categorizes where and how memory is represented, while the conceptual-motivation layer clarifies why particular distinctions matter. Conceptual motivation is not a separate storage technology; instead, it provides lightweight distinctions such as episodic versus semantic memory, working versus long-term memory, procedural memory, consolidation, forgetting, context binding, and prediction-error-driven updating. Nonparametric memory stores information outside model parameters and is usually interpretable, updateable, and modular, but depends heavily on construction and retrieval quality. Parametric memory stores or accesses information through internal model states, trainable modules, KV caches, adapter-based parameterizations, or weights; it can be tightly integrated with inference but is harder to inspect and edit. Memory for agent evolution stores interaction traces, rewards, failures, workflows, code traces, and other experience artifacts so that agents can improve their future behavior. Evaluation is not a memory substrate; it is the empirical layer used to determine whether these designs produce reliable long-term behavior rather than superficial recall.
To make the comparison consistent across sections, Table 2 lists the recurring analytical questions used to compare memory families. These questions shift the review from method enumeration toward lifecycle analysis.
| Dimension | Nonparametric memory | Parametric memory | Agent-evolution memory | Key design question |
| Memory unit | Text records, embeddings, triples, events, media anchors | Hidden states, KV pairs, adapter deltas, trainable vectors, edited weights | Trajectories, rewards, failures, workflows, skills, code traces | What exactly is treated as a reusable memory object? |
| Write policy | Extraction, summarization, graph construction, database update | Cache construction, gradient update, adapter-based adaptation, model editing | Reflection, reinforcement learning, skill abstraction, harness state capture | Who decides what should be stored, and under what confidence threshold? |
| Persistence | Often cross-session and inspectable | Ranges from transient session caches to durable parameter edits | Cross-task or lifelong, but usually tied to agent runtime | Does the memory survive beyond the current conversation or task? |
| Update and forgetting | Explicit add, revise, merge, delete, decay, versioning | Harder to localize; may require cache eviction, re-editing, or retraining | Experience replay, policy revision, skill replacement, trace pruning | Can obsolete or harmful memory be corrected without side effects? |
| Retrieval and use | Similarity search, graph traversal, temporal filtering, symbolic query | Attention reuse, learned lookup, internal activation or weight access | Case retrieval, procedure reuse, context folding, tool-state exposure | Is memory merely recalled, or does it causally improve later behavior? |
| Governance | Strongest support for provenance, permissions, and deletion logs | Weakest support for auditability and deletion | Requires control over experience reuse and self-modification | How can the system prove what it remembered and remove it when needed? |
Table 2. Cross-cutting comparison dimensions for memory mechanisms. The same questions are used throughout the survey to move beyond method-by-method enumeration.
This conceptual layer is not included as a fourth computational mechanism in Table 2; instead, it defines the conceptual memory distinctions used when asking these lifecycle questions across external, internal, and agent-evolution systems.

Figure 3. Detailed mechanism-level taxonomy split from the original full tree. Representative systems and citations are provided in Table 4, Table 6, and Table 7, avoiding an oversized citation-heavy figure in the main text.
| Branch | Memory representation | Lifecycle emphasis | Representative systems |
| Text memory | Facts, events, summaries, user profiles, chunks, embeddings | Selective writing, compression, retrieval, conflict checking, update/delete operations | MemoryBank, Mem0, RF-Mem, Text2Mem, MemoRAG |
| Graph memory | Triples, entities, events, relation paths, temporal or hierarchical graphs | Graph construction, traversal, temporal filtering, multi-hop retrieval, graph maintenance | HippoRAG, RET-LLM, Graphiti/Zep, AriGraph, MAGMA |
| Multimodal memory | Frames, captions, audio transcripts, event segments, story or motion states | Perceptual compression, timestamped retrieval, grounding, identity and temporal consistency | DeepImageSearch, PersonaVLM, VideoAgent, VideoARM, GCAgent, StreamMeCo |
| Memory systems and databases | External stores, user profiles, metadata, indexes, access policies, memory resources | End-to-end lifecycle management, scheduling, provenance, permissions, and governance | Letta/MemGPT, MemOS, MIRIX, LangMem, OpenMemory |
Table 4. Classification of nonparametric memory methods according to this survey.
| Branch | Internal memory object | Main design focus | Representative methods |
| Activation and KV construction | Hidden states, attention keys and values, memory tokens | Extend usable context through recurrence, cached attention, or explicit activation storage | Transformer-XL, Memorizing Transformer, LongMem, MA-LMM, Memory |
| KV-cache management | KV blocks, token caches, compressed attention states | Retain, evict, merge, page, or compress caches while preserving useful context | HO, SnapKV, PyramidKV, ChunkKV, PagedAttention, MemServe |
| Weight editing and adapter-based adaptation | Feed-forward weights, low-rank adapter deltas, side memories | Update durable internal facts or task knowledge with locality and rollback control | ROME, MEMIT, LoRA, WISE, MemoryLLM |
| Learned and compressive memory modules | Trainable memory vectors, virtual memory tokens, neural lookup modules, decoders, sparse experts | Couple memory with model computation through learned lookup, compression, or test-time update | RMem, Engram, Memory Decoder, GradMem, STEM |
Table 6. Classification of parametric memory methods.
| Branch | Stored experience | Evolution role | Representative methods |
| RL and continual learning | Trajectories, rewards, failures, reflections, task outcomes | Use remembered outcomes to improve exploration, planning, self-correction, and policy behavior | MemAgent, Memory-R1, Mem-α, Memento, ReMe, MemRL |
| Procedural and skill memory | Skills, workflows, tool-use patterns, reasoning templates | Abstract successful or failed traces into reusable procedures, workflows, and skills | MemSkill, ReasoningBank, Memento-Skills, AgentEvolver, FLEX |
| Context and harness engineering | Context summaries, files, traces, code, tool protocols, external state | Expose, fold, prune, or externalize runtime state so long workflows remain recoverable | ACON, AgentFold, SWE-Pruner, LCM, MemCompiler |
Table 7. Classification of memory for agent evolution.
Conceptual Motivation from Cognitive Memory
This section is deliberately brief because cognitive-science-inspired memory is used here as a source of terminology and design intuition, not as a separate technical mechanism or a primary contribution of the survey. Figure 2 therefore places it outside the three computational memory families. The relevant value of cognitive memory theory is to sharpen questions that every LLM memory system must answer: whether a stored item is an event, a fact, a temporary task state, or a reusable procedure; whether it should remain local to the current session or become long-term; and whether it should be consolidated, revised, or forgotten.
Human memory research motivates this vocabulary without directly prescribing an LLM architecture. Episodic and semantic distinctions help separate event-level traces from abstracted knowledge [143, 39]; working-memory accounts motivate transient state management [2]; and studies of context binding, neural state spaces, and population codes motivate storing metadata such as time, source, scope, and retrieval conditions rather than storing decontextualized facts alone [10, 58, 40]. These concepts are useful because many LLM-memory failures arise from collapsing different memory types into a single vector store or prompt summary. Table 3 makes this role explicit by mapping cognitive concepts to LLM-memory design implications and the corresponding risks of over-analogy.
| Cognitive concept | LLM-memory design implication | Risk of over-analogy |
| Episodic memory | Store event-level traces with time, participants, source, and task context. | Retrieved episodes are engineered records, not human recollections. |
| Semantic memory | Consolidate repeated observations into abstract facts, preferences, or world knowledge. | Abstraction may erase provenance, uncertainty, and conflicting versions. |
| Working memory | Maintain transient task state, active goals, and short-lived intermediate reasoning. | A long context window or KV cache should not be confused with durable memory. |
| Procedural memory | Preserve reusable workflows, tool-use patterns, and skills for later tasks. | Procedures may encode brittle shortcuts or unsafe habits if not validated. |
| Consolidation and forgetting | Decide which records become durable, which are compressed, and which should decay or be deleted, especially in lifespan-oriented systems and offline recurrence settings ([158, 81]). | Compression and deletion are policy choices, not direct analogues of biological consolidation or sleep. |
| Context binding and prediction-error updating | Attach memories to source, time, scope, confidence, and surprise signals so that updates respond to conflict rather than similarity alone. | Prediction-error metaphors can hide the need for explicit verification, audit logs, and rollback. |
Table 3. Cognitive-memory concepts as design vocabulary for LLM memory. These concepts motivate memory-system questions but should not be interpreted as evidence that LLM memory is cognitively equivalent to human memory.
The same caution applies to consolidation and updating. Work on prediction-error updating and hippocampal–retrosplenial communication suggests that durable memory should be affected by surprise, conflict, and context, not only by semantic similarity [55, 54, 117]. In this survey, however, such findings are treated as conceptual motivation for lifecycle design rather than evidence that current LLM memories are cognitively equivalent to human memory. The core technical analysis begins with the computational substrates below: nonparametric memory, parametric memory, and memory for agent evolution.
Nonparametric Memory
Nonparametric memory stores information outside the model parameters. It is the most visible family of LLM memory methods because it supports direct inspection, deletion, correction, and cross-session persistence. In this section, we review text memory, graph memory, and multimodal memory. Text memory emphasizes compact natural-language records; graph memory emphasizes structured relations and temporal organization; multimodal memory extends memory to visual, audio, video, screen, and embodied observations. Figure 4 illustrates these representative external memory substrates, including database-backed memory systems that manage metadata, indexing, permissions, and persistent stores.
Figure 4. Representative nonparametric LLM memory substrates. External memory may be represented as text chunks and embedding vectors, relational or temporal graphs, multimodal event stores, or system/database layers with controllers, metadata, indexes, permissions, and persistent storage.
Text Memory
Text memory systems transform conversations, documents, tool traces, or observations into natural-language memory units. Their appeal is practical: text records are easy to inspect, edit, embed, and inject back into prompts. The main weakness is also structural: once events, relations, images, or tool states are collapsed into text, retrieval may recover a plausible summary rather than the evidence needed to verify, update, or delete a memory.
Memory Construction and Consolidation
Construction-oriented systems differ less in the fact that they summarize history than in their write policy. MemoryBank [222] and COMEDY [25] show the value of compact conversational summaries, whereas Mem0 [31], RecMem [35], SimpleMem [100], and Light-Mem [45] put more emphasis on salience, delayed consolidation, recursive compression, or staged short-/long-term updates. Personalized and temporal-dialogue systems such as context- and time-sensitive long-term memory [3], Reflective Memory Management [138], MemWeaver [199], and Adaptive Memory Distillation [105] further show that memory construction must decide not only how to summarize but also what deserves to become persistent, hierarchical, or reflective memory. The critical distinction is whether memory writing is eager, selective, recurrent, or sleep-like. Eager writing maximizes recall but accumulates noise and private data; delayed or selective writing reduces cost and noise but can miss rare facts that become important later. A strong text-memory system therefore needs explicit criteria for salience, conflict, source attribution, and deletion, not only a better summarizer.
Memory Retrieval and Update
Retrieval-oriented work shifts the bottleneck from “what to store” to “which stored item should affect the next answer”. Classical sparse, dense, and reranking pipelines, including BM25, DPR Karpukhin et al. [75], and BERT-style rerankers Robertson and Zaragoza [129]; Gao et al. [50], provide useful retrieval baselines, while Text2Mem Wang et al. [160], RF-Mem Zhang et al. [217], MemSearch-o1 Zhang et al. [213], O-Mem Wang et al. [150], MemoRAG Qian et al. [124], and Omne-R1 Liu et al. [98] make retrieval more memory-specific by adding operation languages, uncertainty-aware switching, reasoning-guided search, global memory-enhanced retrieval, or memory-based multi-hop reasoning. The comparison shows that nearest-neighbor similarity is only one part of memory use: long-term agents also need temporal filters, contradiction checks, update decisions, and abstention when memory is weak or stale. Without those controls, a high-recall retriever can amplify outdated or hallucinated memories.
Memory Systems and Databases
End-to-end memory platforms expose memory as an engineering interface rather than as a single algorithm. Letta/MemGPT Packer et al. [122] emphasizes paging between limited context and external recall or archival storage; MemOS Li et al. [92] treats memory as a schedulable resource with metadata and lifecycle operations; MIRIX Wang and Chen [156], LangMem docs LangChain Team [80], Zep Rasmussen et al. [126], MemU repo NevaMind AI [1], and OpenMemory repo Mem0 AI [1] illustrate the shift toward production memory APIs, user profiles, provenance, and cross-session state. These systems are valuable because they make memory auditable and maintainable, but their claims should be separated from benchmarked algorithmic gains: documentation-driven systems often demonstrate interfaces and deployment patterns, whereas peer-reviewed methods more often isolate retrieval, update, or consolidation mechanisms.
Graph Memory
Graph memory externalizes information into nodes, edges, events, and communities. Compared with flat text memory, graphs make relations and temporal links more inspectable, but they also introduce a new dependency: the memory is only as reliable as the entity extraction, relation typing, schema evolution, and graph-maintenance policy that produced it. Knowledge graphs represent entities and relations as interconnected nodes and edges [57, 66]; in LLM memory, this representation is attractive because many long-horizon questions require tracing who did what, when, why, and under which constraints.
Triple- and KG-based systems such as HippoRAG Jiménez Gutiérrez et al. [71], HippoRAG2 Jiménez Gutiérrez et al. [72], RET-LLM Modarressi et al. [115], KGT Sun et al. [136], Path Pooling Wang et al. [147], and PlugMem Yang et al. [190] mainly improve memory by making relational evidence retrievable. Their advantages are strongest when the task depends on multi-hop relations, personalized facts, or explainable paths. Their vulnerabilities are different from text memory: a single wrong edge can mislead several downstream paths, while aggressive graph completion can turn uncertain inferences into durable facts. This makes confidence scores, provenance, temporal validity, and edge-level deletion more important than in ordinary chunk retrieval.
Hierarchical and temporal graph methods, including Graphiti/Zep Rasmussen et al. [126], AriGraph Anokhin et al. [5], MemGAS Xu et al. [181], HyperMem Yue et al. [200], Mnemis Tang et al. [140], MAGMA Jiang et al. [69], and TraceMem Shu et al. [132], move beyond static triples by organizing memory at multiple granularities and across time. This direction is particularly relevant for agents because repeated interactions create evolving communities, goals, tasks, and causal chains rather than isolated facts. The trade-off is maintenance cost: hierarchical graph memories require decisions about when to merge nodes, split episodes, retire obsolete relations, and preserve conflicting versions. Additional work on GraphRAG-style retrieval, graph readers, synthetic KGs, hierarchical generation memory, trainable graph memory, and graph-structured personal assistants [41, 88, 32, 196, 176, 216, 61] is best interpreted through this lens: the key question is not whether a graph exists, but whether graph operations provide more reliable memory use than simpler text retrieval under update pressure.
Multimodal Memory
Multimodal memory extends memory mechanisms to visual, audio, video, screen, and embodied experiences. Its central challenge is deciding what should remain as raw evidence, what should be captioned or summarized, and what should be indexed as an event. If all perceptual history is textualized, retrieval becomes cheap but grounding becomes fragile; if raw media is retained, the system preserves evidence but faces high storage, indexing, and privacy costs.
Image and Visual Memory
Image and visual memory focuses on retaining static or sparsely changing perceptual evidence, such as image histories, screenshots, object appearances, visual preferences, and GUI observations. Compared with video memory, its main difficulty is less about dense temporal compression and more about grounding: the system must preserve object identity, spatial relations, regions, and source images so that later answers are not based only on lossy captions. DeepImageSearch [36] evaluates this setting through context-aware image retrieval over visual histories, where an agent must recover relevant images rather than merely recall text summaries. PersonaVLM [119] highlights personalized visual memory, storing user-specific multimodal preferences and histories so that long-term personalization can use visual evidence instead of generic user-profile text alone. Related screen and GUI memory systems, such as MGA [29], treat interface observations as reusable visual state, requiring memory to preserve widgets, layouts, and interaction traces. Across these image-centered settings, the key design question is whether the memory keeps enough perceptual anchors—images, regions, screenshots, metadata, or embeddings—to support verification and correction after retrieval.
Video and Temporal Multimodal Memory
Video and temporal multimodal memory adds dense time, event boundaries, causal order, object trajectories, audio, and long-horizon consistency to the visual-memory problem. For understanding tasks, systems can be compared by their evidence boundary and write policy. Caption- and event-indexing methods such as Multi-RAG [108], HippoMM [96], VideoAgent [154], and VideoRAG [128] primarily build timestamped textual or event memories from selected frames, audio transcripts, and captions, using retrieval to avoid feeding an entire long video into the model. More explicitly agentic systems expose video memory as a tool interface: VideoAgent [43] stores temporal event descriptions and object-centric tracking states as a structured unified memory; VideoARM [195] couples an observe–think–act–memorize loop with hierarchical multimodal memory for coarse-to-fine long-video reasoning; and GCAgent [193] builds schematic and narrative episodic memories that encode temporal and causal event relations for global-context-aware inference. For embodied settings, Embodied VideoAgent [44] extends video memory with depth, pose, and other sensor-derived scene states, making memory persistent and updateable during interaction. Streaming systems such as HERMES [207], EventMemAgent [166], Visual Agentic Memory [85], and StreamMeCo [148] further shift the focus to online indexing, hierarchical retrieval, KV- or graph-based memory, and compression. These methods show that video memory is not only a question of storing more frames: it requires selective writing, recoverable visual evidence, temporal grounding, and mechanisms for controlling storage and retrieval cost.
For generation and reconstruction, memory becomes a consistency-control mechanism. MotionRAG [225], StoryMem [209], MemFlow [67], MagicWorld [86], Yume [110], and LoGeR [208] use retrieved motion, visual, story, world, or geometric state to prevent drift across long outputs. Related systems such as WorldMM, MemVerse, MGA, M3-Agent, Infinite Video Understanding, long-video memory models, OneStory, Yume-1.5, EgoLCD, Pack-and-Force Memory, Long-Term Spatial Memory, Mixture of Contexts, and Context as Memory (Yeo et al. 2025b; Liu et al. 2025b; Cheng et al. 2025; Long et al. 2025; Zhang et al. 2025a; Wang et al. 2025i; Li et al. 2024b; Weng et al. 2024; Wang et al. 2024d; An et al. 2026; Mao et al. 2025b; Zhang et al. 2025e; Wu et al. 2025d,b; Cai et al. 2025a; Yu et al. 2025b) show that multimodal memory is not only about recall accuracy. It must also be evaluated by grounding fidelity, temporal coherence, identity consistency, and whether the system can trace generated content back to the perceptual or world-state memories that controlled it.
Comparative Discussion
The main strength of nonparametric memory is controllability: memory items can be inspected, edited, deleted, versioned, and connected to explicit sources. However, nonparametric memory also exposes the system to retrieval errors, stale facts, noisy summaries, and inconsistent update rules. The most important design choice is not only which database or index is used, but also the memory unit and write policy. A system that stores every interaction risks noise accumulation and privacy leakage, whereas a system that stores only aggressive summaries may lose evidence needed for later contradiction resolution. Text memory is easiest to deploy but may discard structure. Graph memory supports relation-aware reasoning but depends on reliable extraction and maintenance. Multimodal memory preserves perceptual history but must balance compact captions against access to original visual or audio evidence. Table 5 summarizes these trade-offs.
| Representation | Strengths | Limitations | Suitable scenarios |
| Text memory | Simple, editable, compatible with prompting and retrieval pipelines | Compression may lose temporal, relational, or perceptual details | Dialogue history, user profiles, personal assistants, task notes |
| Graph memory | Preserves entities, relations, events, and multi-hop paths | Extraction noise, schema drift, and costly graph maintenance | Knowledge-intensive reasoning, temporal relation tracking, explainable retrieval |
| Multimodal memory | Maintains links to visual, audio, video, or embodied evidence | High storage cost and difficult grounding between captions and raw media | Video agents, embodied agents, screen agents, long-form generation |
| Database-backed memory | Supports metadata, access control, indexing, and auditability | Requires careful schema design and application integration | Long-running assistants, enterprise agents, multi-user memory systems |
Table 5. Trade-offs among major nonparametric memory representations.
Parametric Memory
Parametric memory refers to mechanisms where knowledge is stored, accessed, or updated through internal model states, trainable modules, weights, or activation caches. This category includes both transient working memory and durable internal memory. Activation states and KV caches usually function as short-term or session-level memory: they extend the effective context and improve inference efficiency, but they do not necessarily persist across independent interactions. Weight edits, trainable memory vectors, and learned modules are closer to long-term parametric memory because they can survive across sessions and alter future behavior, while adapter-based methods are best viewed as parameter-efficient adaptation carriers that become memory-relevant only when tied to a concrete update, composition, or retrieval policy. Compared with nonparametric memory, parametric memory is more tightly coupled with the model and can be efficient during inference, but it is often less transparent and harder to edit safely. Figure 5 shows where these internal memory mechanisms can appear inside a transformer block.
Figure 5. Internal memory mechanisms inside a transformer block. KV caches provide transient session memory around attention, learned memory tokens can be introduced through the input representation, and weight editing or related parameter updates can encode more durable parametric facts in feed-forward or module-level components.
Activation and KV-Cache Memory
Activation and KV-cache methods are often described as memory, but their persistence is usually weaker than that of external stores or weight updates. The KV cache stores keys and values from previous tokens so the model can attend to earlier context without recomputing states; this makes it a form of working memory by default. It becomes long-term memory only when the cache, compressed states, or activation summaries are explicitly stored, retrieved, and reused across sessions.
Construction-oriented methods differ in how they extend the context available to the model. Transformer-XL (Dai et al. [34]) uses segment recurrence, Memorizing Transformer (Wu et al. [174]) augments attention with stored key-value pairs, Long-Mem (Wang et al. [151]) retrieves cached KV memories through an auxiliary network, MA-LMM (He et al. [56]) stores online visual features and learned query representations, and Memory (Yang et al. [189]) frames explicit memory as a storage form beyond parameters and prompt tokens. These methods improve access to prior context, but the memory unit is often an activation state rather than an inspectable fact. As a result, they are strong for continuity and efficiency but weak for provenance, user correction, and deletion.
KV management methods address a different problem: how to keep enough context while reducing inference cost. HO (Zhang et al. [218]), SnapKV (Li et al. [90]), PyramidKV (Cai et al. [19]), ChunkKV (Liu et al. [101]), StreamingLLM (Xiao et al. [178]), FastGen (Ge et al. [51]), MiniCache (Liu et al. [97]), PagedAttention (Kwon et al. [77]), MemServe (Hu et al. [59]), PyramidInfer (Yang et al. [188]), and Infinite-LLM (Lin et al. [93]) vary in whether they select heavy hitters, preserve semantic chunks, compress hierarchically, merge states, or manage cache blocks as a serving resource. The critical evaluation question is therefore not simply throughput or perplexity: a cache policy should also be tested for which memories it silently discards, whether discarded content was causally needed later, and whether compressed cache states can support update or forgetting semantics.
Weight and Module Memory
Weight and module memory is more persistent than activation memory because it changes parameters or parameter-like states that affect future inference. This persistence is valuable when knowledge should be compact and immediately available, but it also makes errors harder to inspect and remove. The main comparison with nonparametric memory is therefore a governance trade-off: internal memory can reduce retrieval overhead, yet it weakens source traceability and precise deletion.
Model-editing methods such as ROME Meng et al. [112] and MEMIT Meng et al. [113] treat selected feed-forward weights as locations for factual associations; WISE Wang et al. [149] separates main and side memory to improve locality and reliability; and MemoryLLM Wang et al. [157] adds a fixed-size latent memory pool inside the transformer. Adapter-based approaches such as LoRA Hu et al. [60] are memory-relevant when the adapter is used as a durable carrier of task, domain, user, or preference information, but standard LoRA is not by itself an online memory system because its updates are normally fixed after training. The key criterion is whether the method specifies a write, composition, retrieval, and rollback policy, not merely whether additional parameters exist.
Learned and compressive modules, including Mem Wang et al. [155], Engram Cheng et al. [30], MLP Memory Geva et al. [53], Memory Decoder Cao et al. [20], GradMem Kuratov et al. [76], MeKi Ding et al. [37], and STEM Sadhukhan et al. [130], further blur the line between neural computation and memory storage. Some reduce explicit retrieval costs by making memories part of the computation graph; others trade retrieval for test-time optimization or sparse lookup. Related work on activation memory, recurrent memory, sparse memory, function-token memory, explicit working memory, infinitely updatable state, memory–reasoning disentanglement, and contextual agentic memory [177, 11, 52, 13, 83, 27, 219, 65, 7, 212, 12, 24, 33, 201, 73, 180]) reinforces the same point: internal memory should be evaluated by persistence, locality, reversibility, interference, and privacy leakage, not only by loss or speed.
Comparative Discussion
Parametric memory is attractive when memory must be close to model computation. KV-cache methods are useful for long-context inference and serving efficiency, but they usually preserve information only for the current or cached session unless explicitly externalized. This makes them different from persistent memory systems even though both are often discussed under the language of memory. Weight editing and learned memory modules can make memories compact and immediately accessible, but they are difficult to audit and delete. Their benefits should therefore be evaluated together with locality, reversibility, interference, and privacy risk. For safety-critical or privacy-sensitive deployments, parametric memory should be paired with external provenance records, edit logs, and evaluation protocols that test whether an internal update changes only the intended behavior rather than silently spreading to unrelated contexts.
Memory for Agent Evolution
Beyond storing information for recall, memory can support agent evolution. In this line of work, memory becomes a substrate for improving behavior: agents store experience, learn from failures, refine procedures, compress successful trajectories, and optimize the context supplied to the LLM. Figure 6 summarizes this self-evolving loop from environment interaction to trajectory storage, reflection, procedural update, and memory-guided task execution.
Figure 6. Self-evolving agent memory loop. Agent memory captures interaction trajectories, retrieves them for reflection and assessment, consolidates insights into procedural or skill memory, and reuses the updated memory to guide future task execution in the environment.
Reinforcement Learning and Continual Learning
Experience-driven memory methods use previous interaction records as training or adaptation signals for future behavior. Their common goal is not merely to recall the past, but to convert past trajectories, errors, rewards, and reflections into better future action. The key design axis is how explicitly the system separates memory operations from task solving. MemAgent [197] trains an agent to update internal memory while processing long contexts; Memory-R1 Yan et al. [186] separates a Memory Manager with ADD, UPDATE, DELETE, and NOOP operations from an Answer Agent; and Mem-α Wang et al. [159] learns to maintain multi-component memories under downstream QA rewards. This separation matters because memory writing becomes an action with consequences, not a passive log.
Other systems emphasize deployment-time adaptation and self-evolution. Memento Zhou et al. [223], ReMe Cao et al. [21], EvolveR Wu et al. [170], MemRL Zhang et al. [214], ProcMEM Mi et al. [114], CASCADE Huang et al. [64], and SEDM Xu et al. [182] use accumulated experiences, cases, procedural abstractions, or distributed self-evolving memories to improve repeated tasks without always fine-tuning the backbone model. Recent work on contextual replay, skill optimization, fast-and-slow learning, long-horizon engineering agents, memory bottlenecks, neural garbage collection, predictive forgetting, test-time learning, nested learning, and task-core memory Liu et al. [102]; Yang et al. [191]; Tiwari et al. [144]; Qiao et al. [125]; Xie et al. [179]; Chen et al. [23]; Wang et al. [162]; Sidik and Rokach [133]; Li et al. [87]; Fountas et al. [47]; Tandon et al. [139]; Liu et al. [103]; Zhang et al. [205]; Yang et al. [187]; Zheng et al. [221]; Yan et al. [184]; Sun et al. [135]; Behrouz et al. [12]; Lan et al. [79]; Lampinen et al. [78]; Zhang et al. [204]; Wu et al. [172]; Lin et al. [95]; Huai et al. [63] broadens the same theme. The unresolved issue is credit assignment: a system must determine which remembered experience caused improvement, which merely correlated with success, and which should be forgotten before it hardens into a brittle strategy.
Procedural and Skill Memory
Procedural memory stores reusable action knowledge rather than factual content. Agent Workflow Memory Wang et al. [165], MemSkill Zhang et al. [206], ReasoningBank Ouyang et al. [121], Memento-Skills Zhou et al. [224], AgentEvolver Zhai et al. [202], FLEX Cai et al. [18], and MUSE-Autoskill Lin et al. [94] all convert successful or failed experience into workflows, memory-operation skills, reasoning templates, policies, or executable procedures. Compared with episodic memory, procedural memory is more compact and easier to reuse across tasks; compared with parametric adaptation, it can remain inspectable as an external artifact.
The risk is that procedures abstract away the evidence that justified them. A workflow that succeeded because of a benchmark-specific shortcut may be stored as a general skill; a reasoning template may preserve a flawed assumption; and a tool-use policy may become unsafe when the environment changes. Procedural-memory systems therefore need skill validation, versioning, retirement, and rollback, not only retrieval. Their evaluation should measure transfer and robustness across tasks rather than repeated success on near-duplicate environments.
Context Engineering and Harness Engineering
A third line treats memory as part of the agent runtime. ACON [74], AgentFold [192], SWE-Pruner [163], LCM [42], the CL-bench repo [6], and the Code as Agent Harness survey [120] emphasize that memory is not only stored content but also the surrounding interface that exposes files, traces, summaries, tools, and execution state. Related work on grep-like harnesses, task-specific memory harnesses, agentic file systems, context engineering, and state-conditioned memory compilation for embodied agents [131, 123, 183, 211, 38] strengthens this point.
The main contribution of this branch is operational rather than mnemonic: it decides which intermediate artifacts remain available and which are hidden, pruned, folded, or externalized. This makes harness memory powerful for software agents, web agents, and long workflows, but also makes evaluation difficult because performance gains may come from better tool exposure rather than better memory construction. A useful benchmark should therefore isolate whether the harness improved memory persistence, retrieval, and update behavior, or merely supplied more convenient context to the base model.
Comparative Discussion
Memory for agent evolution differs from recall-oriented memory because success is measured by behavioral improvement rather than only answer accuracy. RL and continual-learning systems use memory as a feedback substrate; procedural memory abstracts reusable actions; and context and harness engineering preserves the intermediate state needed for reliable execution. These approaches are complementary: an agent may use episodic memories to learn a skill, store the skill as procedural memory, and expose the skill through a harness that manages files, tools, and execution traces. They also differ in risk profile. Experience memories may overfit to spurious successes, procedural memories may preserve brittle or unsafe workflows, and harness-level memories may expose stale files or tool states as if they were reliable evidence. The most important comparison is therefore not which method stores more experience, but which method proves that stored experience causally improves later behavior while remaining reversible, auditable, and safe to retire.
Benchmarks
Evaluating memory-augmented LLMs requires benchmarks that test not only retrieval accuracy but also persistence, temporal reasoning, personalization, multimodal grounding, hallucination control, and long-horizon task execution. Following the organization used in this survey, we divide benchmarks into plain-text benchmarks, multimodal benchmarks, and simulation environments.
Plain-Text Benchmarks
Plain-text benchmarks evaluate memory through conversations, user profiles, documents, and assistant histories. LoCoMo [106] and LoCoMo+ [91] focus on long-term conversational and beyond-factual cognitive memory, while LongMemEval [168] evaluates information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention in long-term chat-assistant histories. BEAM [141], PersonaMem [68], MemoryAgentBench [62], LifelongAgentBench [220], Minerva [175], HaluMem [22], LTM-QA [142], and TENURE [46] emphasize different aspects of memory evaluation, including personalization, lifelong interaction, hallucinated memory use, long-context QA memory, precision-aware retrieval, operation-level hallucination diagnosis, and reliable abstention. Adjacent long-context benchmarks such as Bench [215], LongBench [8], LongBench v2 [9], and NoLiMa [116] are useful controls for separating persistent memory mechanisms from simply placing more evidence inside a longer prompt.
Additional plain-text and text-adjacent benchmarks fill gaps around role-playing histories and memory-specific diagnosis. MOOM [26] evaluates maintenance, organization, and optimization in ultra-long role-playing dialogues; MemBench [137] broadens agent-memory evaluation across memory abilities; and StoryBench [16] is useful as a text-to-video story-continuity benchmark for checking whether generated visual sequences remain consistent across temporally ordered prompts.
Multimodal Benchmarks
Multimodal benchmarks extend memory evaluation beyond text. Video-MME [49] and LVBench [153] test long-video understanding and temporal reasoning. TeleEgo [185] evaluates egocentric multimodal memory, while DeepImageSearch [36] focuses on memory-guided image search. Persona-MME [119] and RoboMemArena [82] further evaluate personalization and robotic or embodied memory. These benchmarks are important because caption-only memory can appear successful on text questions while failing to preserve object identity, spatial relations, or fine-grained perceptual evidence.
The multimodal benchmark landscape also includes long-video and sparse-memory QA resources that are adjacent to agent memory. MovieChat+ [134] tests question-answer sparse memory for long video QA, CinePile [127] provides movie-scale QA, LongVideoBench [169] evaluates interleaved long-context video-language understanding, and EgoSchema [107] diagnoses very long-form egocentric video understanding. These resources help separate genuine multimodal memory from short-window visual recognition.
| Benchmark family | Evaluation target | Memory capability | Task format | Representative benchmarks |
| Plain-text benchmarks | Conversation, user profile, long documents, assistant memory | Multi-session recall, temporal update, abstention, hallucination control, personalization, long-context control | QA, summarization, user modeling, memory judging | LoCoMo<br>LoCoMo+<br>LongMemEval<br>BEAM<br>PersonaMem<br>MemoryAgentBench<br>LifelongAgentBench<br>Minerva<br>HaluMem<br>LTM-QA<br>TENURE<br>∞Bench<br>LongBench<br>LongBench v2<br>NoLiMa |
| Multimodal benchmarks | Video, image, egocentric stream, robotic memory | Visual grounding, long video recall, temporal event tracking, multimodal personalization | Video QA, image search, embodied memory tasks | Video-MME<br>LVBench<br>TeleEgo<br>DeepImageSearch<br>Persona-MME<br>RoboMemArena |
| Simulation environments | Apps, tools, feedback loops, evolving environments | Long-horizon state tracking, task recovery, tool-use memory, procedural reuse | Interactive app tasks, office workflows, simulated conversations | AppWorld<br>MemoryBench<br>AMemGym<br>ARE<br>OdysseyBench |
Table 8. Classification of memory evaluation resources. The table separates evaluation target, memory capability, interaction format, and representative benchmark instead of using method-centered columns.
Simulation Environments
Simulation environments evaluate memory through interactive task completion rather than static question answering. AppWorld [146] builds a controllable world of everyday applications and APIs, requiring agents to remember user instructions, app states, and previous tool results during multi-step tasks. MemoryBench [1] evaluates whether agents can maintain and reuse memory under feedback-driven interactions. AMemGym [70] provides an interactive setting for testing conversational memory over long-horizon assistant sessions. ARE [48] evaluates agents in adaptive or evolving environments where memory must support recovery and continued progress. OdysseyBench [152] focuses on long-horizon office workflows, where agents must coordinate information across documents, spreadsheets, emails, calendars, PDFs, and intermediate files.
Evaluation Metrics and Failure Modes
A benchmark should reveal not only whether the correct answer is produced, but also whether memory was written, updated, retrieved, and used correctly. Table 9 summarizes representative capabilities and failure modes that should be measured across memory systems.
| Capability | What to measure | Typical failure mode | Possible metrics |
| Persistence | Whether important information remains available across sessions | The agent forgets durable facts or long-term preferences | Memory recall, answer accuracy over delayed queries |
| Update and conflict handling | Whether newer evidence revises older memory | The agent retrieves stale or contradictory facts | Update accuracy, contradiction resolution rate |
| Selective writing | Whether the system stores useful information without accumulating noise | Over-writing irrelevant details or under-writing important facts | Memory precision, memory recall, storage growth rate |
| Retrieval grounding | Whether retrieved memory is relevant and correctly used | Irrelevant memory distracts reasoning or produces false continuity | Retrieval precision, citation/evidence correctness, downstream utility |
| Abstention and hallucination control | Whether the model avoids inventing memories | The model claims to remember interactions that never occurred | False-memory rate, abstention accuracy |
| Privacy and deletion | Whether sensitive memory is protected and removable | Private memory leaks across users or persists after deletion | Leakage rate, deletion success rate, access-control violation rate |
Table 9. Capability-oriented evaluation dimensions for memory-augmented LLMs.
Challenges and Future Directions
Unified Memory Organization
Current systems often combine text chunks, embeddings, graphs, KV caches, tool traces, and user profiles in ad hoc ways. Future systems need unified schemas that can align episodic, semantic, procedural, multimodal, and parametric memories without losing provenance or update history. A promising direction is to separate memory content from memory metadata: content stores what is remembered, while metadata records source, timestamp, confidence, permissions, access history, and deletion status. Such organization would make memory systems easier to audit, compare, and integrate across applications.
Reliable Memory Writing
Long-term systems fail when they write too much, write too little, or write incorrect abstractions. Better memory construction should distinguish durable facts, transient preferences, temporary task states, and uncertain observations. It should also support explicit contradiction handling and confidence-aware revision. Future research should move beyond heuristic summarization toward write policies that are evaluated against delayed downstream tasks, not only immediate extraction quality.
Retrieval Beyond Similarity
Nearest-neighbor similarity is insufficient for long-term memory. Useful memories may be old, indirect, private, contradicted, or only relevant under a procedural goal. Retrieval should combine semantic similarity with temporal reasoning, graph traversal, task intent, permission constraints, and uncertainty estimation. An important direction is multi-stage retrieval: first identify the task and memory type needed, then retrieve candidate memories, then verify recency, compatibility, and source reliability before injecting them into the model context.
Multimodal Grounding and Video Intelligence
Multimodal memory must preserve object identity, spatial relations, audiovisual timing, and links to original perceptual evidence. Caption-only memory is compact but can discard important visual details. Future work should balance compact textual summaries with retrievable perceptual anchors such as frames, regions, audio segments, trajectories, or screen states. Evaluation should therefore test whether the system can recover evidence from the original modality rather than only answer questions from textual summaries.
Video-centered intelligence exposes an even broader infrastructure gap. Long-video understanding requires persistent temporal indexes that can track events, entities, causal dependencies, and scene changes across hours of perceptual input, while long-video generation requires memory mechanisms that maintain identity, layout, motion, and narrative consistency across extended outputs. World models and embodied agents further require memory substrates that connect perceptual evidence with latent scene state, physical constraints, action effects, and interaction history. Current systems still often depend on sparse frame sampling, caption-based compression, or short-lived context buffers, which weakens causal tracking and makes generated or embodied behavior drift over long horizons. Future research should therefore develop scalable video memory infrastructure that combines online compression, hierarchical temporal event graphs, multimodal retrieval over raw evidence and latent world states, and evaluation protocols spanning understanding, generation, world modeling, and embodied control.
Governance, Privacy, and Deletion
Long-term memory often contains personal, proprietary, or task-sensitive information. Multi-user, multi-agent, and edge-cloud settings require access control, provenance tracking, permission-aware retrieval, conflict management, secure deletion, and privacy-preserving personalized memory management [28]. Parametric memory is especially challenging because a single remembered fact may not be directly inspectable or removable. Practical memory systems should therefore expose deletion logs, memory scopes, user consent controls, and policy-aware retrieval filters.
These governance mechanisms should be treated as core technical components rather than deployment afterthoughts.
Failure Amplification and Misaligned Adaptation
Memory can amplify errors because stored information influences future actions. A hallucinated fact that is written into memory can become a persistent false belief; an incorrectly extracted graph edge can redirect later multi-hop reasoning; and a noisy user profile can bias personalization across sessions. Agent-evolution memory introduces an additional risk: a workflow that succeeds for the wrong reason may be converted into a reusable skill, causing the agent to repeat brittle strategies or exploit benchmark-specific shortcuts. Future systems should therefore treat memory writing as a high-impact operation rather than a passive logging step. Useful safeguards include uncertainty-aware writes, source-linked memories, conflict detection, periodic memory review, and controlled mechanisms for skill retirement or rollback.
Evaluation Reliability
Current benchmarks cover long-context retrieval, conversational memory, multimodal understanding, and simulation environments, but evaluation remains fragmented. Future benchmarks should measure memory write quality, update correctness, abstention, hallucinated memory use, privacy leakage, temporal consistency, and downstream task success. They should also distinguish between genuine memory use and success through long-context pattern matching. One practical approach is to evaluate agents under controlled memory ablations, where memory contents are removed, corrupted, delayed, or replaced to test whether the model truly depends on the intended memory mechanism.
Conclusion
Memory research for LLMs has moved from simple retrieval augmentation toward a diverse ecosystem of nonparametric stores, graph memories, multimodal memories, parametric memory modules, and self-evolving agent systems. This survey organized the literature around a lightweight cognitive-science-inspired conceptual motivation layer, three computational mechanism families—nonparametric memory, parametric memory, and memory for agent evolution—and a cross-cutting evaluation layer that tests whether these mechanisms produce durable benefits. Across these families, the central design questions are how to construct memory, how to manage and consolidate it, how to retrieve it reliably, how to forget or delete it safely, and how to evaluate its contribution to long-term behavior. The next generation of memory-augmented agents will likely combine text, graph, multimodal, activation, and procedural memories under explicit governance mechanisms, using cognitive concepts mainly as design vocabulary, moving LLMs from reactive generation toward persistent, adaptive, and trustworthy long-term collaboration.
Declarations
Funding
No specific funding was received for this work.
Conflicts of interest/Competing interests
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this manuscript.
Ethics approval
Not applicable. This article is a survey and does not involve human participants, human data, animals, or clinical materials.
Consent to participate
Not applicable.
Consent for publication
Not applicable.
Data availability
This survey is based on publicly available publications, benchmark descriptions, benchmark repositories, system documentation, and project pages. No new dataset was created or analyzed for this manuscript.
Code availability
Not applicable. This article is a survey and does not introduce new code or software.
Authors’ contributions
All authors contributed to the manuscript.
Acknowledgments
The authors thank the creators of the open-source memory systems, benchmarks, and documentation resources that support research on memory-augmented large language model agents.
References
- [1]Ai Q, Tang Y, Wang C, et al. (2025) MemoryBench: A benchmark for memory and continual learning in LLM systems. https://doi.org/10.48550/arXiv.2510.17281, URL https://arxiv.org/abs/2510.17281, arXiv:2510.17281
- [2]Allen RJ, Baddeley AD, Hitch GJ (2026) Awareness as the Heart of Working Memory. Trends in Cognitive Sciences https://doi.org/10.1016/j.tics.2026.04.001, URL https://www.sciencedirect.com/science/article/pii/S1364661326000756
- [3]Alonso N, Figliolia T, Ndirango A, et al (2024) Toward Conversational Agents with Context and Time Sensitive Long-term Memory. https://doi.org/10.48550/arXiv.2406.00057, URL https://arxiv.org/abs/2406.00057, arXiv:2406.00057
- [5]Anokhin P, Semenov N, Sorokin AY, et al (2025) AriGraph: Learning knowledge graph world models with episodic memory for LLM agents. In: Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, pp 12–20, https://doi.org/10.24963/ijcai.2025/2, URL https://doi.org/10.24963/ijcai.2025/2
- [6]Arc Computer (2025) Arc Continual Learning Benchmark: Evaluating LLM agents on stateful environments. URL https://github.com/Arc-Computer/CL-Bench, software benchmark for evaluating LLM agents on continual learning in stateful environments
- [7]Badger BL (2026) Language Model Memory and Memory Models for Language. https://doi.org/10.48550/arXiv.2602.13466, URL https://arxiv.org/abs/2602.13466, arXiv:2602.13466
- [8]Bai Y, Lv X, Zhang J, et al (2024) LongBench: A bilingual, multitask benchmark for long context understanding. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), URL https://arxiv.org/abs/2308.14508, arXiv:2308.14508
- [9]Bai Y, Tu S, Zhang J, et al (2025) LongBench v2: Towards deeper understanding and reasoning on realistic long-context multitasks. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp 3639–3664, URL https://aclanthology.org/2025.acl-long.183/ , arXiv:2412.15204arxiv.org/abs/2412.15204
- [10]Bausch M, Niediek J, Reber TP, et al (2026) Distinct Neuronal Populations in the Human Brain Combine Content and Context. Nature 650:690–700. https://doi.org/10.1038/s41586-025-09910-2, URL https://www.nature.com/articles/s41586-025-09910-2
- [11]Behrouz A, Zhong P, Mirrokni V (2024) Titans: Learning to Memorize at Test Time. https://doi.org/10.48550/arXiv.2501.00663, URL https://arxiv.org/abs/2501.00663, neurIPS 2025 poster, arXiv:2501.00663
- [12]Behrouz A, Razaviyayn M, Zhong P, et al (2025) Nested Learning: The Illusion of Deep Learning Architectures. NeurIPS 2025 poster, URL https://openreview.net/forum?id=nbMeRvNb7A
- [13]Behrouz A, Li Z, Deng Y, et al (2026) Memory Caching: RNNs with Growing Memory. https://doi.org/10.48550/arXiv.2602.24281, URL https://arxiv.org/abs/2602.24281, arXiv:2602.24281
- [14]Brown T, Mann B, Ryder N, et al (2020) Language models are few-shot learners. In: Advances in Neural Information Processing Systems, pp 1877–1901arxiv.org/abs/2005.14165
- [15]Bubeck S, Chandrasekaran V, Eldan R, et al (2023) Sparks of artificial general intelligence: Early experiments with GPT-4. arXiv preprint arXiv:230312712 https://doi.org/10.48550/arXiv.2303.12712, URL https://arxiv.org/abs/2303.12712, arXiv:2303.12712
- [16]Bugliarello E, Moraldo H, Villegas R, et al (2023) StoryBench: A Multifaceted Benchmark for Continuous Story Visualization. In: Advances in Neural Information Processing Systems, URL https://proceedings.neurips.cc/paper_files/paper/2023/hash/f63f5bede4fe08c857c5f377b5d33a-Abstract-Datasets_and_Benchmarks.htmlarxiv.org/abs/2308.11606
- [18]Cai Z, Guo X, Pei Y, et al (2025b) FLEX: Continuous agent evolution via forward learning from experience. https://doi.org/10.48550/arXiv.2511.06449, arXiv:2511.06449
- [19]Cai Z, Zhang Y, Gao B, et al (2025c) PyramidKV: Dynamic KV cache compression based on pyramidal information funneling. In: International Conference on Learning Representations, URL https://openreview.net/forum?id=EBQGK9OE3karxiv.org/abs/2406.02069
- [20]Cao J, Wang J, Wei R, et al (2025) Memory decoder: A pretrained, plug-and-play memory for large language models. arXiv:2508.09874arxiv.org/abs/2508.09874
- [21]Cao Z, Deng J, Yu L, et al (2026) Remember me, refine me: A dynamic procedural memory framework for experience-driven agent evolution. In: Findings of the Association for Computational Linguistics: ACL 2026, https://doi.org/10.48550/arXiv.2512.10696, to appear, arXiv:2512.10696
- [22]Chen D, Niu S, Li K, et al (2025a) HaluMem: Evaluating hallucinations in memory systems of agents. URL https://arxiv.org/abs/2511.03506, arXiv:2511.03506
- [23]Chen G, Chen J, Chen L, et al (2026a) Toward Autonomous Long-Horizon Engineering for ML Research. https://doi.org/10.48550/arXiv.2604.13018, URL https://arxiv.org/abs/2604.13018, arXiv:2604.13018
- [24]Chen M, Li Y, Padhte K, et al (2025b) Improving Factuality with Explicit Working Memory. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Vienna, Austria, pp 11999–11213, https://doi.org/10.18653/v1/2025.acl-long.548, URL https://aclanthology.org/2025.acl-long.548/
- [25]Chen N, Li H, Chang J, et al (2025c) Compress to impress: Unleashing the potential of compressive memory in real-world long-term conversations. In: Proceedings of the 31st International Conference on Computational Linguistics. Association for Computational Linguistics, pp 755–773, URL https://aclanthology.org/2025.coling-main.51/arxiv.org/abs/2402.11975
- [26]Chen W, Tang J, Hou Z, et al (2025d) MOOM: Maintenance, Organization and Optimization of Memory in Ultra-Long Role-Playing Dialogues. https://doi.org/10.48550/arXiv.2509.11860, URL https://arxiv.org/abs/2509.11860, arXiv:2509.11860
- [27]Chen Y, Chen R, Yi S, et al (2026b) MSA: Memory Sparse Attention for Efficient End-to-End Memory Model Scaling to 100M Tokens. https://doi.org/10.48550/arXiv.2603.23516, URL https://arxiv.org/abs/2603.23516, arXiv:2603.23516
- [28]Chen Y, Zhao J, Tang B, et al (2026c) MemPrivacy: Privacy-Preserving Personalized Memory Management for Edge-Cloud Agents. https://doi.org/10.48550/arXiv.2605.09530, URL https://arxiv.org/abs/2605.09530, arXiv:2605.09530
- [29]Cheng W, Liu J, Sun Y, et al (2025) MGA: Memory-Driven GUI Agent for Observation-Centric Interaction. https://doi.org/10.48550/arXiv.2510.24168, URL https://arxiv.org/abs/2510.24168, arXiv:2510.24168
- [30]Cheng X, Zeng W, Dai D, et al (2026) Conditional memory via scalable lookup: A new axis of sparsity for large language models. arXiv:2601.07372arxiv.org/abs/2601.07372
- [31]Chhikara P, Khant D, Aryan S, et al (2025) Mem0: Building production-ready ai agents with scalable long-term memory. arXiv preprint arXiv:250419413 https://doi.org/10.48550/arXiv.2504.19413, URL https://arxiv.org/abs/2504.19413, arXiv:2504.19413
- [32]Choubey P, Su X, Luo M, et al (2024) Distill-synthkg: Distilling knowledge graph synthesis workflow for improved coverage and efficiency. URL https://arxiv.org/abs/2410.16597, arXiv:2410.16597
- [33]Chukwu E, Bindschaedler L (2025) May the Memory Be With You: Efficient and Infinitely Updatable State for Large Language Models. In: The 5th Workshop on Machine Learning and Systems (EuroMLSys 2025). ACM, Rotterdam, Netherlands, https://doi.org/10.1145/3721146.3721951, URL https://dl.acm.org/doi/10.1145/3721146.3721951
- [34]Dai Z, Yang Z, Yang Y, et al (2019) Transformer-XL: Attentive language models beyond a fixed-length context. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, Florence, Italy, pp 2978–2988, https://doi.org/10.18653/v1/P19-1285
- [35]Dai Z, Deng S, Guan S, et al (2026) RecMem: Recurrence-based memory consolidation for efficient and effective long-running llm agents. arXiv preprint arXiv:2605160405
- [36]Deng C, Deng M, Wu J, et al (2026) DeepImageSearch: Benchmarking multimodal agents for context-aware image retrieval in visual histories. URL https://arxiv.org/abs/2602.10809, arXiv:2602.10809
- [37]Ding N, Liu F, Kim K, et al (2026a) MeKi: Memory-based expert knowledge injection for efficient LLM scaling. arXiv:2602.03359arxiv.org/abs/2602.03359
- [38]Ding X, Wang X, Yang Y, et al (2026b) MemCompiler: Compile, Don’t Inject – State-Conditioned Memory for Embodied Agents. https://doi.org/10.48550/arXiv.2605.07594, URL https://arxiv.org/abs/2605.07594, arXiv:2605.07594
- [39]Dong CV, Lu Q, Norman KA, et al (2025) Towards Large Language Models with Human-Like Episodic Memory. Trends in Cognitive Sciences 29(10):928–941. https://doi.org/10.1016/j.tics.2025.06.016, URL https://www.cell.com/trends/cognitive-sciences/abstract/S1364-6613(25)00179-2
- [40]Dupret D, Fusi S, Panzeri S (2025) Neural Population Activity for Memory: Properties, Computations, and Codes. Neuron 114(3):390–407. https://doi.org/10.1016/j.neuron.2025.11.007, URL https://www.cell.com/neuron/fulltext/S0896-6273(25)00854-2
- [41]Edge D, Trinh H, Cheng N, et al (2024) From local to global: A graph RAG approach to query-focused summarization. URL https://arxiv.org/abs/2404.16130, arXiv:2404.16130
- [42]Ehrlich C, Blackman T (2026) LCM: Lossless context management. URL https://arxiv.org/abs/2605.04050, arXiv:2605.04050
- [43]Fan Y, Ma X, Wu R, et al (2024) VideoAgent: A memory-augmented multimodal agent for video understanding. In: European Conference on Computer Vision, URL https://arxiv.org/abs/2403.11481, arXiv:2403.11481
- [44]Fan Y, Ma X, Su R, et al (2025) Embodied VideoAgent: Persistent memory from egocentric videos and embodied sensors enables dynamic scene understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, URL https://arxiv.org/abs/2501.00358, arXiv:2501.00358
- [45]Fang J, Deng X, Xu H, et al (2025) Lightmem: Lightweight and efficient memory-augmented generation. arXiv preprint arXiv:251018866arxiv.org/abs/2510.18866
- [46]Flynt J (2026) Structured Belief State and the First Precision-Aware Benchmark for LLM Memory Retrieval. https://doi.org/10.48550/arXiv.2605.11325, URL https://arxiv.org/abs/2605.11325, arXiv:2605.11325
- [47]Fountas Z, Oomerjee A, Bou-Ammar H, et al (2026) Why the Brain Consolidates: Predictive Forgetting for Optimal Generalisation. https://doi.org/10.48550/arXiv.2603.04688, URL https://arxiv.org/abs/2603.04688, arXiv:2603.04688
- [48]Froger R, Andrews P, Bettini M, et al (2025) ARE: Scaling up agent environments and evaluations. https://doi.org/10.48550/arXiv.2509.17158, URL https://arxiv.org/abs/2509.17158, arXiv:2509.17158
- [49]Fu C, Dai Y, Luo Y, et al (2025) Video-MME: The first-ever comprehensive evaluation benchmark of multi-modal LLMs in video analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 24108–24118, URL https://arxiv.org/abs/2405.21075, arXiv:2405.21075
- [50]Gao L, Dai Z, Callan J (2021) Rethink training of BERT rerankers in multi-stage retrieval pipeline. In: Advances in Information Retrieval: 43rd European Conference on IR Research, ECIR 2021, Proceedings, Part II, Lecture Notes in Computer Science, vol 12657. Springer, pp 280–286, https://doi.org/10.1007/978-3-030-72240-1_26, URL https://doi.org/10.1007/978-3-030-72240-1_26
- [51]Ge S, Zhang Y, Liu L, et al (2024) Model tells you what to discard: Adaptive KV cache compression for LLMs. In: International Conference on Learning Representations, URL https://arxiv.org/abs/2310.01801, arXiv:2310.01801
- [52]Gelberg Y, Eitan Y, Bronstein M, et al (2026) Training Transformers for KV Cache Compressibility. https://doi.org/10.48550/arXiv.2605.05971, URL https://arxiv.org/abs/2605.05971, arXiv:2605.05971
- [53]Geva M, Schuster R, Berant J, et al (2021) Transformer feed-forward layers are key-value memories. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic, pp 5484–5495, https://doi.org/10.18653/v1/2021.emnlp-main.446
- [54]Gonzalez J, V⁷⁷or”oslakos M, Aykan D, et al (2026) Subspace Communication in the Hippocampal–Retrosplenial Axis. Nature https://doi.org/10.1038/s41586-026-10481-z, URL https://www.nature.com/articles/s41586-026-10481-z
- [55]Groves I, Grella SL, Harley CW, et al (2025) How Prediction Error Drives Memory Updating: Role of Locus Coeruleus–Hippocampal Interactions. Trends in Neurosciences 48(11). https://doi.org/10.1016/j.tins.2025.09.003, URL https://www.cell.com/trends/neurosciences/abstract/S0166-2236(25)00189-4
- [56]He B, Li H, Jang YK, et al (2024) MA-LMM: Memory-augmented large multimodal model for long-term video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 13504–13514, https://doi.org/10.1109/CVPR52733.2024.01282
- [57]Hogan A, Blomqvist E, Cochez M, et al (2021) Knowledge graphs. ACM Computing Surveys 54(4):71:1–71:37. https://doi.org/10.1145/3447772, URL https://doi.org/10.1145/3447772
- [58]Horner AJ (2026) A Neural State Space for Episodic Memories. Trends in Cognitive Sciences https://doi.org/10.1016/j.tics.2025.10.009, URL https://www.cell.com/trends/cognitive-sciences/fulltext/S1364-6613(25)00284-0
- [59]Hu C, Huang H, Hu J, et al (2024) MemServe: Context caching for disaggregated LLM serving with elastic memory pool. URL https://arxiv.org/abs/2406.17565, arXiv:2406.17565
- [60]Hu EJ, Shen Y, Wallis P, et al (2022) LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations, URL https://openreview.net/forum?id=nZeVKeeFYf9arxiv.org/abs/2106.09685
- [61]Hu M, Chen T, Chen Q, et al (2025a) HiAgent: Hierarchical Working Memory Management for Solving Long-Horizon Agent Tasks with Large Language Model. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Vienna, Austria, pp 32779–32798, https://doi.org/10.18653/v1/2025.acl-long.1575, URL https://aclanthology.org/2025.acl-long.1575/
- [62]Hu Y, Wang Y, McAuley J (2025b) Evaluating memory in LLM agents via incremental multi-turn interactions. URL https://arxiv.org/abs/2507.05257, arXiv:2507.05257
- [63]Huai T, Zhou J, Cai Y, et al (2025) Task-Core Memory Management and Consolidation for Long-term Continual Learning. https://doi.org/10.48550/arXiv.2505.09952, URL https://arxiv.org/abs/2505.09952, arXiv:2505.09952
- [64]Huang X, Chen J, Fei Y, et al (2025) CASCADE: Cumulative agentic skill creation through autonomous development and evolution. https://doi.org/10.48550/arXiv.2512.23880, arXiv:2512.23880
- [65]Huang Z, Min Q, Huang H, et al (2024) Ultra-Sparse Memory Network. https://doi.org/10.48550/arXiv.2411.12364, URL https://arxiv.org/abs/2411.12364, ICLR 2025 poster, arXiv:2411.12364
- [66]Ji S, Pan S, Cambria E, et al (2022) A survey on knowledge graphs: Representation, acquisition, and applications. IEEE Transactions on Neural Networks and Learning Systems 33(2):494–514. https://doi.org/10.1109/TNNLS.2021.3070843, URL https://doi.org/10.1109/TNNLS.2021.3070843
- [67]Ji S, Chen X, Yang S, et al (2025) MemFlow: Flowing adaptive memory for consistent and efficient long video narratives. arXiv preprint arXiv:251214699arxiv.org/abs/2512.14699
- [68]Jiang B, Hao Z, Cho YM, et al (2025) Know me, respond to me: Benchmarking LLMs for dynamic user profiling and personalized responses at scale. In: Second Conference on Language Modeling, https://doi.org/10.48550/arXiv.2504.14225, URL https://openreview.net/forum?id=6ox8XZGOpP, arXiv:2504.14225
- [69]Jiang D, Li Y, Li G, et al (2026) Magma: A multi-graph based agentic memory architecture for ai agents. arXiv preprint arXiv:260103236
- [70]Jiayang C, Ru D, Qiu L, et al (2026) AMemGym: Interactive memory benchmarking for assistants in long-horizon conversations. URL https://arxiv.org/abs/2603.01966, arXiv:2603.01966
- [71]Jiménez Gutiérrez B, Shu Y, Gu Y, et al (2024) HippoRAG: Neurobiologically inspired long-term memory for large language models. In: Advances in Neural Information Processing Systems, URL https://arxiv.org/abs/2405.14831, arXiv:2405.14831
- [72]Jiménez Gutiérrez B, Shu Y, Qi W, et al (2025) From RAG to memory: Non-parametric continual learning for large language models. In: Proceedings of the International Conference on Machine Learning, URL https://arxiv.org/abs/2502.14802, arXiv:2502.14802
- [73]Jin M, Luo W, Cheng S, et al (2025) Disentangling Memory and Reasoning Ability in Large Language Models. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Vienna, Austria, pp 1681–1701, https://doi.org/10.18653/v1/2025.acl-long.84, URL https://aclanthology.org/2025.acl-long.84/
- [74]Kang M, Chen WN, Han D, et al (2026) ACON: Optimizing context compression for long-horizon LLM agents. In: Proceedings of the 43rd International Conference on Machine Learning, https://doi.org/10.48550/arXiv.2510.00615, to appear, arXiv:2510.00615
- [75]Karpukhin V, Oguz B, Min S, et al (2020) Dense passage retrieval for open-domain question answering. In: Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), pp 6769–6781arxiv.org/abs/2004.04906
- [76]Kuratov Y, Kairov M, Bulatov A, et al (2026) GradMem: Learning to write context into memory with test-time gradient descent. URL https://arxiv.org/abs/2603.13875, arXiv:2603.13875
- [77]Kwon W, Li Z, Zhuang S, et al (2023) Efficient memory management for large language model serving with PagedAttention. In: Proceedings of the 29th Symposium on Operating Systems Principles. Association for Computing Machinery, pp 611–626, https://doi.org/10.1145/3600006.3613165
- [78]Lampinen AK, Engelcke M, Li Y, et al (2025) Latent Learning: Episodic Memory Complements Parametric Learning by Enabling Flexible Reuse of Experiences. https://doi.org/10.48550/arXiv.2509.16189, URL https://arxiv.org/abs/2509.16189, arXiv:2509.16189
- [79]Lan H, Yu Y, Qian L, et al (2025) LightSearcher: Efficient DeepSearch via Experiential Memory. https://doi.org/10.48550/arXiv.2512.06653, URL https://arxiv.org/abs/2512.06653, arXiv:2512.06653
- [80]LangChain Team (2025) LangMem: LangChain’s memory primitives for storing, recalling, and managing agent state in langgraph workflows. Software documentation and blog post, URL https://blog.langchain.com/langmem-sdk-launch/
- [81]Lee S, McLeish S, Goldstein T, et al (2026) Do Language Models Need Sleep? Offline Recurrence for Improved Online Inference. https://doi.org/10.48550/arXiv.2605.26099, URL https://arxiv.org/abs/2605.26099, arXiv:2605.26099
- [82]Lei H, Song W, Zhang H, et al (2026a) RoboMemArena: A comprehensive and challenging robotic memory benchmark. URL https://arxiv.org/abs/2605.10921, arXiv:2605.10921
- [83]Lei J, Zhang D, Li J, et al (2026b) δ-mem: Efficient Online Memory for Large Language Models. https://doi.org/10.48550/arXiv.2605.12357, URL https://arxiv.org/abs/2605.12357, arXiv:2605.12357
- [84]Lewis P, Perez E, Piktus A, et al (2020) Retrieval-augmented generation for knowledge-intensive NLP tasks. In: Advances in Neural Information Processing Systems, pp 9459–9474arxiv.org/abs/2005.11401
- [85]Li AY, Numan N, Steed A (2026a) Visual agentic memory: Enabling online long video understanding via online indexing, hierarchical memory, and agentic retrieval. arXiv preprint arXiv:260516481
- [86]Li G, Li B, Chen J, et al (2025a) MagicWorld: Towards long-horizon stability for interactive video world exploration. arXiv preprint arXiv:251118886
- [87]Li MY, Hamid JI, Fox EB, et al (2026b) Neural Garbage Collection: Learning to Forget while Learning to Reason. https://doi.org/10.48550/arXiv.2604.18002, URL https://arxiv.org/abs/2604.18002, arXiv:2604.18002
- [88]Li S, He Y, Guo H, et al (2024a) GraphReader: Building graph-based agent to enhance long-context abilities of large language models. URL https://arxiv.org/abs/2406.14550, arXiv:2406.14550
- [90]Li Y, Huang Y, Yang B, et al (2024c) SnapKV: LLM knows what you are looking for before generation. In: Advances in Neural Information Processing Systems, URL https://arxiv.org/abs/2404.14469, arXiv:2404.14469
- [91]Li Y, Guo W, Zhang L, et al (2026c) Locomo-Plus: Beyond-factual cognitive memory evaluation framework for LLM agents. https://doi.org/10.48550/arXiv.2602.10715, URL https://arxiv.org/abs/2602.10715, arXiv:2602.10715
- [92]Li Z, Xi C, Li C, et al (2025b) MemOS: A memory OS for AI system. arXiv preprint arXiv:2507.03724 https://doi.org/10.48550/arXiv.2507.03724, URL https://arxiv.org/abs/2507.03724, arXiv:2507.03724
- [93]Lin B, Peng T, Zhang C, et al (2024) Infinite-LLM: Efficient LLM service for long context with disttattention and distributed KVCache. URL https://arxiv.org/abs/2401.02669, arXiv:2401.02669
- [94]Lin H, Li P, Song J, et al (2026) MUSE-Autoskill: Self-evolving agents via skill creation, memory, management, and evaluation. URL https://arxiv.org/abs/2605.27366, arXiv:2605.27366
- [95]Lin J, Zettlemoyer L, Ghosh G, et al (2025a) Continual Learning via Sparse Memory Finetuning. https://doi.org/10.48550/arXiv.2510.15103, URL https://arxiv.org/abs/2510.15103, arXiv:2510.15103
- [96]Lin Y, Zhang J, Wang Q, et al (2025b) HippoMM: Hippocampal-inspired multimodal memory for long audiovisual event understanding. URL https://arxiv.org/abs/2504.10739, arXiv:2504.10739
- [97]Liu A, Liu J, Pan Z, et al (2024) MiniCache: KV cache compression in depth dimension for large language models. In: Advances in Neural Information Processing Systems, vol 37. Curran Associates, Inc., pp 139997–140031, https://doi.org/10.52202/079017-4443
- [98]Liu B, Ji F, Nan J, et al (2025a) Omne-R1: Learning to Reason with Memory for Multi-hop Question Answering. https://doi.org/10.48550/arXiv.2508.17330, URL https://arxiv.org/abs/2508.17330, arXiv:2508.17330
- [100]Liu J, Su Y, Xia P, et al (2026a) Simplemem: Efficient lifelong memory for llm agents. arXiv preprint arXiv:260102553
- [101]Liu X, Tang Z, Dong P, et al (2026b) Chunkkv: Semantic-preserving kv cache compression for efficient long-context llm inference. pp 28728–28778arxiv.org/abs/2502.00299
- [102]Liu Y, Si C, Narasimhan K, et al (2025c) Contextual Experience Replay for Self-Improvement of Language Agents. https://doi.org/10.48550/arXiv.2506.06698, URL https://arxiv.org/abs/2506.06698, arXiv:2506.06698
- [103]Liu Z, Cai Y, Zhu X, et al (2025d) ML-Master: Towards AI-for-AI via Integration of Exploration and Reasoning. https://doi.org/10.48550/arXiv.2506.16499, URL https://arxiv.org/abs/2506.16499, arXiv:2506.16499
- [105]Ma W, Nan J, Wu W, et al (2025) What Deserves Memory: Adaptive Memory Distillation for LLM Agents. https://doi.org/10.48550/arXiv.2508.03341, URL https://arxiv.org/abs/2508.03341, arXiv:2508.03341
- [106]Maharana A, Lee DH, Tulyakov S, et al (2024) Evaluating very long-term conversational memory of LLM agents. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Bangkok, Thailand, pp 13851–13870, https://doi.org/10.18653/v1/2024.acl-long.747, URL https://aclanthology.org/2024.acl-long.747/
- [107]Mangalam K, Akshulakov R, Malik J (2023) EgoSchema: A Diagnostic Benchmark for Very Long-form Video Language Understanding. In: Advances in Neural Information Processing Systems, URL https://proceedings.neurips.cc/paper_files/paper/2023/hash/90ce332aff156b910b002ce4e688bb80-Abstract-Datasets-and-Benchmarks.htmlarxiv.org/abs/2308.09126
- [108]Mao M, Perez-Cabarcas MM, Kallakuri U, et al (2025a) Multi-rag: A multimodal retrieval-augmented generation system for adaptive video understanding. arXiv preprint arXiv:250523990arxiv.org/abs/2505.23990
- [110]Mao X, Lin S, Li Z, et al (2025c) Yume: An interactive world generation model. arXiv preprint arXiv:250717744arxiv.org/abs/2507.17744
- [112]Meng K, Bau D, Andonian A, et al (2022) Locating and editing factual associations in GPT. In: Advances in Neural Information Processing Systems, vol 35. Curran Associates, Inc., pp 17359–17372arxiv.org/abs/2202.05262
- [113]Meng K, Sen Sharma A, Andonian A, et al (2023) Mass-editing memory in a transformer. In: International Conference on Learning Representations, URL https://openreview.net/forum?id=MkbcAHIYgSarxiv.org/abs/2210.07229
- [114]Mi Q, Ma Z, Yang M, et al (2026) Procerm: Learning reusable procedural memory from experience via non-parametric ppo for llm agents
- [115]Modarressi A, Imani A, Fayyaz M, et al (2023) RET-LLM: Towards a general read-write memory for large language models. URL https://arxiv.org/abs/2305.14322, arXiv:2305.14322
- [116]Modarressi A, Deilamsalehy H, Dernoncourt F, et al (2025) NoLiMa: Long-context evaluation beyond literal matching. URL https://arxiv.org/abs/2502.05167, arXiv:2502.05167
- [117]Morrow E, Huang R, Clewett DV (2026) Dopaminergic Processes Predict Temporal Distortions in Event Memory. Nature Communications 17(1). https://doi.org/10.1038/s41467-026-69950-8, URL https://www.nature.com/articles/s41467-026-69950-8
- [120]Ning X, Tieu K, Fu D, et al (2026) Code as agent harness. https://doi.org/10.48550/arXiv.2605.18747, URL https://arxiv.org/abs/2605.18747, arXiv:2605.18747
- [121]Ouyang S, Yan J, Hsu IH, et al (2026) ReasoningBank: Scaling agent self-evolving with reasoning memory. In: International Conference on Learning Representations, URL https://openreview.net/forum?id=jL7fWchScmarxiv.org/abs/2509.25140
- [122]Packer C, Wooders S, Lin K, et al (2023) MemGPT: Towards LLMs as operating systems. https://doi.org/10.48550/arXiv.2310.08560, URL https://arxiv.org/abs/2310.08560, arXiv:2310.08560
- [123]Pan W, Liu S, Zhou X, et al (2026) M*: Every Task Deserves Its Own Memory Harness. https://doi.org/10.48550/arXiv.2604.11811, URL https://arxiv.org/abs/2604.11811, arXiv:2604.11811
- [124]Qian H, Liu Z, Zhang P, et al (2025) MemorAG: Boosting Long Context Processing with Global Memory-Enhanced Retrieval Augmentation. https://doi.org/10.48550/arXiv.2409.05591, URL https://arxiv.org/abs/2409.05591, arXiv:2409.05591
- [125]Qiao J, Meng W, Cheng Y, et al (2026) Memory Intelligence Agent. https://doi.org/10.48550/arXiv.2604.04503, URL https://arxiv.org/abs/2604.04503, arXiv:2604.04503
- [126]Rasmussen P, Paliychuk P, Beauvais T, et al (2025) Zep: a temporal knowledge graph architecture for agent memory. arXiv preprint arXiv:250113956arxiv.org/abs/2501.13956
- [127]Rawal R, Saifullah K, Farr’e M, et al (2024) CinePile: A Long Video Question Answering Dataset and Benchmark. https://doi.org/10.48550/arXiv.2405.08813, URL https://arxiv.org/abs/2405.08813, arXiv:2405.08813
- [128]Ren X, Xu L, Xia L, et al (2026) Videorag: Retrieval-augmented generation with extreme long-context videos. In: Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, pp 2390–2401arxiv.org/abs/2502.01549
- [129]Robertson S, Zaragoza H (2009) The probabilistic relevance framework: BM25 and beyond, vol 4. Now Publishers IncDOI
- [130]Sadhukhan R, Cao S, Dong H, et al (2026) STEM: Scaling transformers with embedding modules. In: International Conference on Learning Representations, URL https://openreview.net/forum?id=gufRimweSQ
- [131]Sen S, Kasturi A, Lumer E, et al (2026) Is Grep All You Need? How Agent Harnesses Reshape Agentic Search. https://doi.org/10.48550/arXiv.2605.15184, URL https://arxiv.org/abs/2605.15184, arXiv:2605.15184
- [132]Shu Y, Liu P, Zhang T, et al (2026) Tracemem: Weaving narrative memory schemata from user conversational traces. arXiv preprint arXiv:260209712
- [133]Sidik B, Rokach L (2026) MEMTIER: Tiered Memory Architecture and Retrieval Bottleneck Analysis for Long-Running Autonomous AI Agents. https://doi.org/10.48550/arXiv.2605.03675, URL https://arxiv.org/abs/2605.03675, arXiv:2605.03675
- [134]Song E, Chai W, Ye T, et al (2024) MovieChat+: Question-aware Sparse Memory for Long Video Question Answering. https://doi.org/10.48550/arXiv.2404.17176, URL https://arxiv.org/abs/2404.17176, arXiv:2404.17176
- [135]Sun C, Chen X, Luo J, et al (2025) Beyond Heuristics: A Decision-Theoretic Framework for Agent Memory Management. https://doi.org/10.48550/arXiv.2512.21567, URL https://arxiv.org/abs/2512.21567, arXiv:2512.21567
- [136]Sun J, Du Z, Chen Y (2024) Knowledge graph tuning: Real-time large language model personalization based on human feedbackarxiv.org/abs/2405.19686
- [137]Tan H, Zhang Z, Ma C, et al (2025a) MemBench: Towards More Comprehensive Evaluation on the Memory of LLM-based Agents. https://doi.org/10.48550/arXiv.2506.21605, URL https://arxiv.org/abs/2506.21605, arXiv:2506.21605
- [138]Tan Z, Yan J, Hsu IH, et al (2025b) In Prospect and Retrospect: Reflective Memory Management for Long-term Personalized Dialogue Agents. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Vienna, Austria, pp 8416–8439, https://doi.org/10.18653/v1/2025.acl-long.413, URL https://aclanthology.org/2025.acl-long.413/
- [139]Tandon A, Dalal K, Li X, et al (2025) End-to-End Test-Time Training for Long Context. https://doi.org/10.48550/arXiv.2512.23675, URL https://arxiv.org/abs/2512.23675, arXiv:2512.23675
- [140]Tang Z, Yu X, Xiao Z, et al (2026) Mnemis: Dual-route retrieval on hierarchical graphs for long-term llm memory. arXiv preprint arXiv:260215313
- [141]Tavakoli M, Salemi A, Ye C, et al (2026) Beyond a million tokens: Benchmarking and enhancing long-term memory in LLMs. In: International Conference on Learning Representations, URL https://arxiv.org/abs/2510.27246, arXiv:2510.27246
- [142]Terranova A, Ross B, Birch A (2025) Evaluating Long-Term Memory for Long-Context Question Answering. https://doi.org/10.48550/arXiv.2510.23730, URL https://arxiv.org/abs/2510.23730, arXiv:2510.23730
- [143]Tibon R, Greve A, Humphreys G, et al (2026) Neural Activations and Representations during Episodic versus Semantic Memory Retrieval. Nature Human Behaviour 10(4):803–821. https://doi.org/10.1038/s41562-025-02390-4, URL https://www.nature.com/articles/s41562-025-02390-4
- [144]Tiwari R, Sareen K, Agrawal LA, et al (2026) Learning, Fast and Slow: Towards LLMs That Adapt Continually. https://doi.org/10.48550/arXiv.2605.12484, URL https://arxiv.org/abs/2605.12484, arXiv:2605.12484
- [145]Touvron H, Lavril T, Izacard G, et al (2023) LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:230213971 https://doi.org/10.48550/arXiv.2302.13971, URL https://arxiv.org/abs/2302.13971, arXiv:2302.13971
- [146]Trivedi H, Khot T, Hartmann M, et al (2024) AppWorld: A controllable world of apps and people for benchmarking interactive coding agents. URL https://arxiv.org/abs/2407.18901, arXiv:2407.18901
- [147]Wang H, Feng Y, Xie X, et al (2025a) Path pooling: Training-free structure enhancement for efficient knowledge graph retrieval-augmented generation. arXiv preprint arXiv:250305203arxiv.org/abs/2503.05203
- [148]Wang J, Sun T, Zhu J, et al (2026a) StreamMeCo: Long-term agent memory compression for efficient streaming video understanding. URL https://arxiv.org/abs/2604.09000, arXiv:2604.09000
- [149]Wang P, Li Z, Zhang N, et al (2024a) WISE: Rethinking the knowledge memory for lifelong model editing of large language models. In: Advances in Neural Information Processing Systems, URL https://arxiv.org/abs/2405.14768, arXiv:2405.14768
- [150]Wang P, Tian M, Li J, et al (2025b) O-mem: Omni memory system for personalized, long horizon, self-evolving agents. arXiv preprint arXiv:251113593arxiv.org/abs/2511.13593
- [151]Wang W, Dong L, Cheng H, et al (2023) Augmenting language models with long-term memory. In: Proceedings of the 37th International Conference on Neural Information Processing Systems. Curran Associates Inc., Red Hook, NY, USA, NIPS ’23arxiv.org/abs/2306.07174
- [152]Wang W, Han D, Diaz DM, et al (2025c) OdysseyBench: Evaluating LLM agents on long-horizon complex office application workflows. URL https://arxiv.org/abs/2508.09124, arXiv:2508.09124
- [153]Wang W, He Z, Hong W, et al (2025d) LVBench: An extreme long video understanding benchmark. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, URL https://arxiv.org/abs/2406.08035, arXiv:2406.08035
- [154]Wang X, Zhang Y, Zohar O, et al (2024b) VideoAgent: Long-form video understanding with large language model as agent. In: European Conference on Computer Vision, URL https://videoagent.github.io/, arXiv:2403.10517arxiv.org/abs/2403.10517
- [155]Wang X, Wang S, Zhu Y, et al (2025e) R³Mem: Bridging memory retention and retrieval via reversible compression. In: Findings of the Association for Computational Linguistics: ACL 2025. Association for Computational Linguistics, Vienna, Austria, pp 4541–4557, https://doi.org/10.18653/v1/2025.findings-acl.235
- [156]Wang Y, Chen X (2025) Mirix: Multi-agent memory system for llm-based agents. arXiv preprint arXiv:250707957arxiv.org/abs/2507.07957
- [157]Wang Y, Gao Y, Chen X, et al (2024c) MEMORYLLM: Towards self-updatable large language models. In: Proceedings of the International Conference on Machine Learning, URL https://arxiv.org/abs/2402.04624, arXiv:2402.04624
- [158]Wang Y, Han C, Wu T, et al (2025f) Towards LifeSpan Cognitive Systems. https://doi.org/10.48550/arXiv.2409.13265, URL https://arxiv.org/abs/2409.13265, published in Transactions on Machine Learning Research, 2025, arXiv:2409.13265
- [159]Wang Y, Takanobu R, Liang Z, et al (2025g) Mem-α: Learning memory construction via reinforcement learning. arXiv preprint arXiv:250925911arxiv.org/abs/2509.25911
- [160]Wang Y, Yang L, Chen B, et al (2025h) Text2mem: A unified memory operation language for memory operating system. arXiv preprint arXiv:250911145arxiv.org/abs/2509.11145
- [162]Wang Y, Chen X, Jin X, et al (2026b) OpenClaw-RL: Train Any Agent Simply by Talking. https://doi.org/10.48550/arXiv.2603.10165, URL https://arxiv.org/abs/2603.10165, arXiv:2603.10165
- [163]Wang Y, Shi Y, Yang M, et al (2026c) SWE-Pruner: Self-adaptive context pruning for coding agents. URL https://arxiv.org/abs/2601.16746, arXiv:2601.16746
- [165]Wang ZZ, Mao J, Fried D, et al (2025j) Agent workflow memory. In: Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research, vol 267. PMLR, pp 63897–63911, URL https://proceedings.mlr.press/v267/wang25bx.htmlarxiv.org/abs/2409.07429
- [166]Wen S, Wang Z, Zhang X, et al (2026) Eventmemagent: Hierarchical event-centric memory for online video understanding with adaptive tool use. arXiv preprint arXiv:260215329
- [168]Wu D, Wang H, Yu W, et al (2024a) LongMemEval: Benchmarking chat assistants on long-term interactive memory. URL https://arxiv.org/abs/2410.10813, arXiv:2410.10813
- [169]Wu H, Li D, Chen B, et al (2024b) LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding. https://doi.org/10.48550/arXiv.2407.15754, URL https://arxiv.org/abs/2407.15754, arXiv:2407.15754
- [170]Wu R, Wang X, Mei J, et al (2025a) EvolveR: Self-evolving LLM agents through an experience-driven lifecycle. URL https://arxiv.org/abs/2510.16079, arXiv:2510.16079
- [172]Wu X, Li K, Zhao Y, et al (2025c) ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization. https://doi.org/10.48550/arXiv.2509.13313, URL https://arxiv.org/abs/2509.13313, arXiv:2509.13313
- [174]Wu Y, Rabe MN, Hutchins D, et al (2022) Memorizing transformers. In: International Conference on Learning Representationsarxiv.org/abs/2203.08913
- [175]Xia M, Rühle V, Rajmohan S, et al (2025a) Minerva: A programmable memory test benchmark for language models. In: Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research, vol 267. PMLR, pp 68258–68276, URL https://proceedings.mlr.press/v267/xia25c.htmlarxiv.org/abs/2502.03358
- [176]Xia S, Xu Z, Chai J, et al (2025b) From Experience to Strategy: Empowering LLM Agents with Trainable Graph Memory. https://doi.org/10.48550/arXiv.2511.07800, URL https://arxiv.org/abs/2511.07800, arXiv:2511.07800
- [177]Xiao C, Zhang P, Han X, et al (2024a) InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory. https://doi.org/10.48550/arXiv.2402.04617, URL https://arxiv.org/abs/2402.04617, arXiv:2402.04617
- [178]Xiao G, Tian Y, Chen B, et al (2024b) Efficient streaming language models with attention sinks. In: International Conference on Learning Representations, URL https://openreview.net/forum?id=NG7sS51zVFarxiv.org/abs/2309.17453
- [179]Xie Z, Hu Z, Ye F, et al (2026) PASK: Toward Intent-Aware Proactive Agents with Long-Term Memory. https://doi.org/10.48550/arXiv.2604.08000, URL https://arxiv.org/abs/2604.08000, arXiv:2604.08000
- [180]Xu B, Dai X, Zhang K (2026) Contextual Agentic Memory is a Memo, Not True Memory. https://doi.org/10.48550/arXiv.2604.27707, URL https://arxiv.org/abs/2604.27707, arXiv:2604.27707
- [181]Xu D, Wen Y, Jia P, et al (2025a) From single to multi-granularity: Toward long-term memory association and selection of conversational agents. arXiv preprint arXiv:250519549arxiv.org/abs/2505.19549
- [182]Xu H, Hu J, Zhang K, et al (2025b) SEDM: Scalable Self-Evolving Distributed Memory for Agents. https://doi.org/10.48550/arXiv.2509.09498, URL https://arxiv.org/abs/2509.09498, arXiv:2509.09498
- [183]Xu X, Mao R, Bai Q, et al (2025c) Everything is Context: Agentic File System Abstraction for Context Engineering. https://doi.org/10.48550/arXiv.2512.05470, URL https://arxiv.org/abs/2512.05470, arXiv:2512.05470
- [184]Yan B, Li C, Qian H, et al (2025a) General Agentic Memory Via Deep Research. https://doi.org/10.48550/arXiv.2511.18423, URL https://arxiv.org/abs/2511.18423, arXiv:2511.18423
- [185]Yan J, Ren R, Liu J, et al (2025b) TeleEgo: Benchmarking egocentric AI assistants in the wild. https://doi.org/10.48550/arXiv.2510.23981, URL https://arxiv.org/abs/2510.23981, arXiv:2510.23981
- [186]Yan S, Yang X, Huang Z, et al (2025c) Memory-R1: Enhancing large language model agents to manage and utilize memories via reinforcement learning. URL https://arxiv.org/abs/2508.19828, arXiv:2508.19828
- [187]Yang C, Yang X, Wen L, et al (2025) Learning on the Job: An Experience-Driven Self-Evolving Agent for Long-Horizon Tasks. https://doi.org/10.48550/arXiv.2510.08002, URL https://arxiv.org/abs/2510.08002, arXiv:2510.08002
- [188]Yang D, Han X, Gao Y, et al (2024a) PyramidInfer: Pyramid KV cache compression for high-throughput LLM inference. In: Findings of the Association for Computational Linguistics: ACL 2024. Association for Computational Linguistics, Bangkok, Thailand, pp 3258–3270, https://doi.org/10.18653/v1/2024.findings-acl.195
- [189]Yang H, Lin Z, Wang W, et al (2024b) Memory³: Language modeling with explicit memory. URL https://arxiv.org/abs/2407.01178, arXiv:2407.01178
- [190]Yang K, Chen Z, He X, et al (2026a) PlugMem: A task-agnostic plugin memory module for LLM agents. https://doi.org/10.48550/arXiv.2603.03296, URL https://arxiv.org/abs/2603.03296, arXiv:2603.03296
- [191]Yang Y, Gong Z, Huang W, et al (2026b) SkillOpt: Executive Strategy for Self-Evolving Agent Skills. https://doi.org/10.48550/arXiv.2605.23904, URL https://arxiv.org/abs/2605.23904, arXiv:2605.23904
- [192]Ye R, Zhang Z, Li K, et al (2026) AgentFold: Long-horizon web agents with proactive context folding. In: International Conference on Learning Representations, URL https://openreview.net/forum?id=IuZoTgsUwsarxiv.org/abs/2510.24699
- [193]Yeo JH, Chung S, Park S, et al (2025a) Gcagent: Long-video understanding via schematic and narrative episodic memory. URL https://arxiv.org/abs/2511.12027, arXiv:2511.12027
- [195]Yin Y, Meng Q, Chen M, et al (2026) VideoARM: Agentic reasoning over hierarchical memory for long-form video understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, URL https://arxiv.org/abs/2512.12360, arXiv:2512.12360
- [196]Yu EJ, Zhu D, Song Y, et al (2025a) Hierarchical Memory Organization for Wikipedia Generation. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Vienna, Austria, https://doi.org/10.18653/v1/2025.acl-long.1423, URL https://aclanthology.org/2025.acl-long.1423/
- [197]Yu H, Chen T, Feng J, et al (2026) MemAgent: Reshaping long-context LLM with multi-conv RL-based memory agent. In: International Conference on Learning Representations, URL https://openreview.net/forum?id=k5nIOvYGCLarxiv.org/abs/2507.02259
- [199]Yu S, Cheng M, Wang D, et al (2025c) MemWeaver: A Hierarchical Memory from Textual Interactive Behaviors for Personalized Generation. https://doi.org/10.48550/arXiv.2510.07713, URL https://arxiv.org/abs/2510.07713, arXiv:2510.07713
- [200]Yue J, Hu C, Sheng J, et al (2026) Hyperm em: Hypergraph memory for long-term conversations. arXiv preprint arXiv:260408256
- [201]Zanzotto FM, Ruzzetti ES, Xompero GA, et al (2025) Position Paper: MeMo: Towards Language Models with Associative Memory Mechanisms. In: Findings of the Association for Computational Linguistics: ACL 2025. Association for Computational Linguistics, Vienna, Austria, pp 15169–15180, https://doi.org/10.18653/v1/2025.findings-acl.785, URL https://aclanthology.org/2025.findings-acl.785/
- [202]Zhai Y, Tao S, Chen C, et al (2025) AgentEvolver: Towards efficient self-evolving agent system. https://doi.org/10.48550/arXiv.2511.10395, arXiv:2511.10395
- [204]Zhang G, Fu M, Yan S (2025b) MemGen: Weaving Generative Latent Memory for Self-Evolving Agents. https://doi.org/10.48550/arXiv.2509.24704, URL https://arxiv.org/abs/2509.24704, arXiv:2509.24704
- [205]Zhang G, Ren H, Zhan C, et al (2025c) MemEvolve: Meta-Evolution of Agent Memory Systems. https://doi.org/10.48550/arXiv.2512.18746, URL https://arxiv.org/abs/2512.18746, arXiv:2512.18746
- [206]Zhang H, Long Q, Bao J, et al (2026a) MemSkill: Learning and evolving memory skills for self-evolving agents. URL https://arxiv.org/abs/2602.02474, arXiv:2602.02474
- [207]Zhang H, Yang S, Fu J, et al (2026b) Hermes: Kv cache as hierarchical memory for efficient streaming video understanding. arXiv preprint arXiv:260114724
- [208]Zhang J, Herrmann C, Hur J, et al (2026c) LoGeR: Long-context geometric reconstruction with hybrid memory. arXiv preprint arXiv:260303269
- [209]Zhang K, Jiang L, Wang A, et al (2025d) StoryMem: Multi-shot long video storytelling with memory. arXiv preprint arXiv:251219539
- [211]Zhang Q, Hu C, Upasani S, et al (2025f) Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models. https://doi.org/10.48550/arXiv.2510.04618, URL https://arxiv.org/abs/2510.04618, arXiv:2510.04618
- [212]Zhang S, Lin Y, Li H (2025g) Memory Retrieval and Consolidation in Large Language Models through Function Tokens. https://doi.org/10.48550/arXiv.2510.08203, URL https://arxiv.org/abs/2510.08203, arXiv:2510.08203
- [213]Zhang S, Li J, Zhang Y, et al (2026d) Memsearch-o1: Empowering large language models with reasoning-aligned memory growth in agentic search. arXiv preprint arXiv:260417265
- [214]Zhang S, Wang J, Zhou R, et al (2026e) MemRL: Self-evolving agents via runtime reinforcement learning on episodic memory. URL https://arxiv.org/abs/2601.03192, arXiv:2601.03192
- [215]Zhang X, Chen Y, Hu S, et al (2024) ∞bench: Extending long context evaluation beyond 100k tokens. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp 15262–15277, URL https://aclanthology.org/2024.acl-long.814/, arXiv:2402.13718arxiv.org/abs/2402.13718
- [216]Zhang Y, Yuan W, Jiang Z (2025h) Bridging Intuitive Associations and Deliberate Recall: Empowering LLM Personal Assistant with Graph-Structured Long-term Memory. In: Findings of the Association for Computational Linguistics: ACL 2025. Association for Computational Linguistics, Vienna, Austria, pp 17533–17547, https://doi.org/10.18653/v1/2025.findings-acl.901, URL https://aclanthology.org/2025.findings-acl.901/
- [217]Zhang Y, Li J, Zhang W, et al (2026f) Evoking user memory: Personalizing llm via recollection-familiarity adaptive retrieval
- [218]Zhang Z, Sheng Y, Zhou T, et al (2023) H2O: Heavy-hitter oracle for efficient generative inference of large language models. In: Advances in Neural Information Processing Systems, URL https://openreview.net/forum?id=RkRrPp7GKODOI
- [219]Zhao T, Jones L (2026) Fast-weight Product Key Memory. https://doi.org/10.48550/arXiv.2601.00671, URL https://arxiv.org/abs/2601.00671, arXiv:2601.00671
- [220]Zheng J, Cai X, Li Q, et al (2025a) LifelongAgentBench: Evaluating LLM agents as lifelong learners. URL https://arxiv.org/abs/2505.11942, arXiv:2505.11942
- [221]Zheng Y, McKee KL, Miconi T, et al (2025b) Goal-Directed Search Outperforms Goal-Agnostic Memory Compression in Long-Context Memory Tasks. https://doi.org/10.48550/arXiv.2511.21726, URL https://arxiv.org/abs/2511.21726, arXiv:2511.21726
- [222]Zhong W, Guo L, Gao Q, et al (2024) Memorybank: Enhancing large language models with long-term memory. In: Proceedings of the AAAI Conference on Artificial Intelligence, pp 19724–19731, https://doi.org/10.1609/aaai.v38i17.29946, URL https://doi.org/10.1609/aaai.v38i17.29946
- [223]Zhou H, Chen Y, Guo S, et al (2025) Memento: Fine-tuning LLM agents without fine-tuning LLMs. https://doi.org/10.48550/arXiv.2508.16153, arXiv:2508.16153
- [224]Zhou H, Guo S, Liu A, et al (2026) Memento-skills: Let agents design agents. https://doi.org/10.48550/arXiv.2603.18743, arXiv:2603.18743
- [225]Zhu C, Wu Y, Wang S, et al (2025) Motion retrieval-augmented image-to-video generation. URL https://arxiv.org/abs/2509.26391, arXiv:2509.26391



