PaperlayerView exact revision on arXiv

Source-linked AI summary

Memory in the Age of AI Agents

Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang, Boyang Liu, Fangyi Zhu, Jiahang Lin, Honglin Guo, Shihan Dou, Zhiheng Xi, Senjie Jin, Jiejun Tan, Yanbin Yin, Jiongnan Liu, Zeyu Zhang, Zhongxiang Sun, Yutao Zhu, Hao Sun, Boci Peng, Zhenrong Cheng, Xuanbo Fan, Jiaxin Guo, Xinlei Yu, Zhenhong Zhou, Zewen Hu, Jiahao Huo, Junhao Wang, Yuwei Niu, Yu Wang, Zhenfei Yin, Xiaobin Hu, Yue Liao, Qiankun Li, Kun Wang, Wangchunshu Zhou, Yixin Liu, Dawei Cheng, Qi Zhang, Tao Gui, Shirui Pan, Yan Zhang, Philip Torr, Zhicheng Dou, Ji-Rong Wen, Xuanjing Huang, Yu-Gang Jiang, Shuicheng Yan

arXiv:2512.13564v2cs.CLcs.AI

TL;DR

Agent memory research spans fragmented mechanisms and remains limited by passive context handling, verbatim retrieval, and unclear update behavior. This survey organizes the field through forms, functions, and dynamics, concluding that memory is essential for temporal coherence, continual adaptation, and long-horizon competence while identifying learnable, adaptive, self-organizing systems as emerging directions.

  • Problem

    Current agent memory systems often treat context as a passive buffer and rely on verbatim retrieval, limiting active workspace management and constructive memory dynamics.

  • Method

    The survey organizes agent memory through the unified lenses of forms, functions, and dynamics, including token-level, parametric, and latent realizations and factual, experiential, and working functions.

  • Results

    The survey concludes that memory is an essential substrate for agents’ temporal coherence, continual adaptation, and long-horizon competence.

  • Takeaways & Limitations

    Emerging directions integrate reinforcement learning with multimodal, multi-agent, and generative memory toward fully learnable, adaptive, and self-organizing systems.

  • Takeaways & Limitations

    Memory updating faces a stability–plasticity dilemma: incorrect updates can overwrite critical information and cause knowledge degradation and faulty reasoning.

Abstract

from arXiv · show

Memory has emerged, and will continue to remain, a core capability of foundation model-based agents. As research on agent memory rapidly expands and attracts unprecedented attention, the field has also become increasingly fragmented. Existing works that fall under the umbrella of agent memory often differ substantially in their motivations, implementations, and evaluation protocols, while the proliferation of loosely defined memory terminologies has further obscured conceptual clarity. Traditional taxonomies such as long/short-term memory have proven insufficient to capture the diversity of contemporary agent memory systems. This work aims to provide an up-to-date landscape of current agent memory research. We begin by clearly delineating the scope of agent memory and distinguishing it from related concepts such as LLM memory, retrieval augmented generation (RAG), and context engineering. We then examine agent memory through the unified lenses of forms, functions, and dynamics. From the perspective of forms, we identify three dominant realizations of agent memory, namely token-level, parametric, and latent memory. From the perspective of functions, we propose a finer-grained taxonomy that distinguishes factual, experiential, and working memory. From the perspective of dynamics, we analyze how memory is formed, evolved, and retrieved over time. To support practical development, we compile a comprehensive summary of memory benchmarks and open-source frameworks. Beyond consolidation, we articulate a forward-looking perspective on emerging research frontiers, including memory automation, reinforcement learning integration, multimodal memory, multi-agent memory, and trustworthiness issues. We hope this survey serves not only as a reference for existing work, but also as a conceptual foundation for rethinking memory as a first-class primitive in the design of future agentic intelligence.

1 Introduction

As LLMs evolve into adaptive agents, memory has become a foundational capability, yet rapid methodological advances and inconsistent terminology make the field difficult to organize. This survey addresses the gap with a multidimensional framework covering agent-memory forms, functions, dynamics, resources, and emerging directions.

  • Motivation: Agent memory supports continual adaptation by transforming static LLMs into agents that evolve through environmental interaction.Its importance spans applications requiring proactive memory management and long-term interaction.
  • Motivation: Existing surveys and taxonomies no longer fully reflect rapid methodological advances in contemporary agent memory research.The survey therefore motivates an updated perspective on the field.
  • Motivation: Diverse memory terminologies, including declarative, episodic, semantic, and parametric memory, further obscure conceptual clarity.The paper calls for a coherent taxonomy that unifies these emerging concepts.
  • Framework: The survey organizes agent memory through forms, functions, and dynamics, addressing its definition, purposes, and operation over time.It distinguishes factual, experiential, and working memory and presents lifecycle dynamics through formulation, retrieval, and evolution.
  • Scope and resources: It additionally compares agent memory with related concepts, summarizes benchmarks and open-source frameworks, and examines emerging research frontiers.Frontiers include automation-oriented design, reinforcement learning, multimodal and multi-agent memory, and trustworthy issues.
  • Contributions: The survey presents an up-to-date multidimensional taxonomy intended to structure understanding of current agent memory developments.It also discusses how memory types can align with distinct agentic objectives and compiles resources for researchers and practitioners.

2 Preliminaries: Formalizing Agents and Memory

The paper formalizes agents as autoregressive decision-makers operating through observations, memory retrieval, computation, and actions, then models memory as an evolving state integrated through formation, evolution, and retrieval. This framework covers single- and multi-agent systems, within-task and cross-task memory, and persistent cognitive support beyond a single context window.

  • 2.1 LLM-based Agent Systems: LLM-based agents are formalized to include both single-agent and multi-agent configurations coupled to memory through read/write interactions.The formulation covers reasoning, tool use, collaboration, and other agent settings.
  • 2.1 LLM-based Agent Systems: Each agent receives an observation history and task specification, then produces actions through an autoregressive LLM policy conditioned on contextual inputs.Visible history can include messages, tool outputs, reasoning traces, shared workspace states, and other agents’ contributions.
  • 2.1 LLM-based Agent Systems: The action space spans natural-language generation, tool invocation, planning, direct environment control, and communication with other agents.These actions support interaction with APIs, software repositories, embodied environments, shared memory buffers, and collaborators.
  • 2.1 LLM-based Agent Systems: Agent execution forms trajectories that interleave environment observation, optional memory retrieval, LLM computation, and action execution across time.The next environment state results from the executed action, while stopping is determined by task or system criteria.
  • 2.2 Agent Memory Systems: Agent memory is modeled as an evolving state that can support both cross-task information retained before a task and short-term information accumulated during it.The state may be implemented as a text buffer, key–value store, vector database, graph, or hybrid representation.
  • 2.2 Agent Memory Systems: Memory dynamics comprise formation from informational artifacts, evolution of stored candidates, and context-dependent retrieval for action selection.Formation selectively extracts potentially reusable information; evolution can consolidate, resolve conflicts, discard low-utility entries, or restructure memory.
  • 2.2 Agent Memory Systems: Short-term and long-term memory arise from the temporal patterns of formation, evolution, and retrieval rather than from discrete architectural modules.Retrieval may occur at initialization, intermittently, or continuously, while memory updates may be episodic or continuous.
  • 2.3.3 Agent Memory vs. Context Engineering: Context engineering manages the momentary interface between agent and model, whereas agent memory sustains a persistent cognitive state beyond a single context window.The comparison distinguishes transient resource management from support for learning, adaptation, and autonomy.

3 Form: What Carries Memory?

The survey organizes agent memory by where information resides and how it is represented, identifying token-level, parametric, and latent forms. This structural framework supports analysis of memory’s functions and dynamics.

  • Agent memory takes different storage forms across task settings rather than one unified structure.These forms have distinct structural properties and capabilities.
  • The survey identifies three memory forms: token-level, parametric, and latent memory.The categories are organized by where memory resides and how it is represented.
  • Token-level Memory: Token-level memory uses explicit, discrete units that remain externally visible and can be accessed, modified, and reconstructed.
  • Parametric Memory: Parametric memory encodes information in model parameters and accesses it implicitly during forward computation.
  • Latent Memory: Latent memory uses hidden states, continuous representations, or evolving latent structures that can persist and update across interactions.
  • The three forms establish a structural framework for examining memory’s functions, evolution, adaptation, and effects on agent behavior.

3.1 Token-level Memory

Token-level memory stores persistent, inspectable discrete units, whose organization ranges from topology-free collections to single-layer and multi-layer structures. The survey relates these forms to retrieval, updating, reasoning, and diverse memory contents and applications.

  • Token-level memory stores persistent, discrete units that are externally accessible and inspectable.Units can include text tokens, visual tokens, audio frames, and other discrete elements stored outside model parameters.
  • Token-level memory is transparent, editable, and interpretable, supporting retrieval, routing, conflict handling, and coordination with other memory forms.It is described as the most common memory form with the largest body of existing work.
  • Flat Memory (1D): Flat Memory (1D) accumulates sequences or bags of units without explicit inter-unit topology.Examples include snippets, trajectories, chunks, text chunks, profiles, experience trajectories, vector representations, and multimodal entries.
  • Planar Memory (2D): Planar Memory (2D) organizes related units within one structured layer using graphs, trees, tables, or similar structures.It has explicit structure but no cross-layer relations.
  • Hierarchical Memory (3D): Hierarchical Memory (3D) structures memory across multiple layers connected by inter-layer links.This organization forms a volumetric or stratified memory.
  • The organizational spectrum governs how token-level memory supports search, updating, reasoning, and capabilities such as abstraction and cross-layer reasoning.Figure 3 presents flat, planar, and hierarchical organization as progressively more topologically complex.
  • Token-level memory research uses diverse unit granularities and structures, including compact semantic representations, query-response pairs, hybrid vector-table systems, and cognitive categories.Some systems also store inductive thoughts, reflections, and decomposed factual, plot-level, or character information.
  • Flat Memory (1D): Flat Memory scales through inexpensive appending, pruning, and similarity retrieval but lacks relational organization for coherent structured inference.As collections grow, redundancy and noise may accumulate, limiting compositional reasoning, long-horizon planning, and abstraction formation.

3.2 Parametric Memory

Parametric memory stores information in model parameters rather than visible token units, and is organized by whether storage resides in the original weights or auxiliary parameters. The section compares these forms, their training phases, applications, and trade-offs.

  • Parametric memory embeds information in learnable parameter spaces, allowing models to internalize and recall it without external storage.
  • Internal parametric memory directly adjusts the original weights, whereas external parametric memory stores information in auxiliary parameters such as adapters or LoRA modules.
  • Parametric memory methods can inject knowledge during pre-training, mid-training, or post-training for knowledge retrieval, agent experience, long-window performance, personalization, or learning from past executions.
  • Internal parameters avoid extra inference overhead and deployment costs but are difficult to update, costly to retrain, and prone to forgetting old memory.
  • External parameter modules support modular updates, task-specific personalization, controlled rollback, and reduced interference with pretrained representations, but their effectiveness depends on integration with internal representations.

3.3 Latent Memory

Latent memory carries information in internal representations rather than readable tokens or dedicated parameter sets. The section classifies it by how latent states are generated, reused, or transformed, emphasizing efficiency alongside information-loss and complexity risks.

  • Latent memory stores KV caches, activations, hidden states, or latent embeddings as implicit internal representations rather than explicit tokens or dedicated parameter sets.
  • Latent representations can capture dense contextual, long-range, and cross-modal signals while avoiding repeated processing of full contexts during extended interactions.
  • Generate creates reusable latent representations with an auxiliary model or module, Reuse carries forward prior computational states, and Transform compresses existing states.
  • Generate: Generated latent memory compresses long contexts, task trajectories, or multimodal inputs into compact states that can be stored, inserted, or reused later.
  • Latent-memory generation and transformation may lose information, introduce bias or accumulated errors, and require additional computation, data, and engineering complexity.
  • Reuse leverages internal activations such as curated KV representations, while Transform reshapes states through selection, retrieval enhancement, or compressed re-encoding.

3.4 Adaptation

The section frames memory-type selection as a design choice tied to expected agent behavior and task requirements. It contrasts token-level transparency, parametric generalization, and latent flexibility and efficiency across application settings.

  • Memory-type selection reflects how designers expect information to shape agent behavior for a given task.
  • Token-level Memory: Token-level memory is symbolic, addressable, transparent, and suited to explicit reasoning, controllability, accountability, and frequent updates.
  • Token-level Memory: Token-level memory aligns with dialogue, lifelong agents, personalization, recommendation, enterprise knowledge, and high-stakes domains requiring verifiable provenance.
  • Parametric Memory: Parametric memory is implicit, abstract, and generalizable, making it suited to broad conceptual understanding, structural insight, and behavioral or domain-specific priors.
  • Latent Memory: Latent memory balances flexibility and efficiency through low readability, privacy protection, expressive semantic encoding, and compact cross-modal representations.
  • Latent Memory: Latent memory is presented as suitable for multimodal architectures, edge or cloud deployment, and encrypted or privacy-sensitive applications.

4 Functions: Why Agents Need Memory?

Agent memory addresses the statelessness of LLM-based agents by supporting persistence, adaptation, coherent interaction, and active context management. The survey organizes its functions into factual, experiential, and working memory.

  • Factual Memory: Factual memory stores explicit facts, user information, and environmental states to support consistency, coherence, and adaptability.It supports context awareness, personalized responses, and extended task planning.
  • Experiential Memory: Experiential memory abstracts past trajectories, failures, and successes into procedural or strategic knowledge for continual learning and self-evolution.The survey also describes adaptation as converting interaction feedback into reusable knowledge that can rectify errors and refine decisions.
  • Working Memory: Working memory is a capacity-limited, dynamically controlled scratchpad for managing active context within a task or session.Its purpose is to make the context window controllable, updatable, and resistant to interference.
  • Experiential Memory: Case-based memory preserves raw trajectories with high fidelity, while strategy-based memory distills generalizable schemas that guide planning but are not executable actions.Cases provide verifiable evidence but can burden retrieval and context capacity; strategies constrain search and improve robustness on unseen tasks.
  • Experiential Memory: Skill-based memory operationalizes abstract insights as callable procedures, APIs, or tools that connect planning with environmental interaction.Robust agents can combine strategies for planning logic with skills for grounded execution.
  • Working Memory: Single-turn working memory compresses static inputs, whereas multi-turn working memory consolidates interaction state to preserve temporal coherence and goal alignment.Plans and structured environment representations can remain readable and writable, supporting robust strategy revision against perception failures.

5 Dynamics: How Memory Operates and Evolves?

Agentic memory is presented as a dynamic system rather than static storage. It extracts and generalizes knowledge from reasoning traces and environmental feedback, then fuses and updates it with existing memory.

  • Dynamics: Dynamic memory management extends beyond fixed parameters or databases by constructing, updating, and retrieving memory conditionally over time.The survey frames this as a shift from static storage to dynamic memory utilization.
  • Memory Evolution: Agents can extract refined, generalizable knowledge from reasoning traces and environmental feedback, then integrate it with the existing memory base.This describes memory as an evolving store rather than a fixed repository.

1. Memory Formulation

The supplied passage identifies knowledge distillation and structured construction as memory-formulation components, alongside latent representation and parametric internalization.

  • Memory Formulation: Memory formulation includes knowledge distillation and structured construction.These labels appear as formulation-related components in the supplied material.
  • Scope: The supplied formulation labels do not specify how these components are implemented or evaluated.Only the named components are stated in the passage.

3. Memory Retrieval

The supplied retrieval material labels query-related memories and existing similar memories, but does not provide a fuller retrieval procedure.

  • Scope: The supplied retrieval passages do not specify the inputs, outputs, or evaluation of the retrieval process.They contain stage labels rather than a complete operational account.
  • Memory Retrieval: Memory retrieval includes querying memories related to the current input or task.The supplied passage presents this as a retrieval-stage label.
  • Memory Retrieval: Existing similar memories are identified as another retrieval-stage element.The passage provides the label without describing its ranking or selection mechanism.

2. Memory Evolution

Agent memory operates through three interconnected processes: formation extracts useful knowledge from experience, evolution integrates it with existing memory, and retrieval accesses task-relevant information. Their feedback cycle supports continuous adaptation to changing environments.

  • Memory formation transforms raw interactive experiences into information-dense knowledge units by selectively identifying patterns with long-term utility.
  • Memory evolution integrates newly formed memories with the existing memory base through consolidation, conflict resolution, and adaptive pruning.
  • Memory retrieval constructs a task-aware query from context and accesses an appropriate memory bank to support reasoning.
  • The three processes form an interconnected cycle in which reasoning outcomes and environmental feedback generate new memories and refine the memory base.

5.1 Memory Formation

Memory formation compresses raw experiences into representations selected for long-term utility, using complementary strategies that differ in granularity, structure, and machine-level encoding. These approaches improve efficiency and capability, but trade off detail, cross-partition coherence, transparency, or update flexibility.

  • 5.1 Memory Formation: Memory formation selectively encodes raw contexts into compact knowledge to address the computational and storage costs of lengthy, noisy, redundant inputs.
  • 5.1 Memory Formation: The survey organizes memory formation by compression granularity and encoding logic, while noting that strategies can be combined within one algorithm.
  • Semantic Summarization: Semantic summarization reduces contextual overhead by preserving global meaning while filtering redundancy from lengthy raw data.
  • Knowledge Distillation: Knowledge distillation extracts cognitive assets such as factual details, planning strategies, and step-level execution insights from experience.
  • Structured Construction / Latent Representation: Structured construction organizes source data into graphs or hierarchical trees to improve explainability and support multi-hop reasoning, while latent representation encodes experiences as vectors or KV states for multimodal alignment.
  • Semantic Summarization: Incremental summarization progressively fuses new information into an evolving summary, avoiding the O(n^2) burden of full-sequence processing but retaining serial-update bottlenecks and forgetting risks.
  • Semantic Summarization: Partitioned summarization processes semantic sub-chunks separately, improving efficiency and fine-grained coverage while risking loss of cross-partition dependencies.
  • Parametric Internalization: Parametric internalization embeds experiential capabilities in model weights for zero-latency access, but faces catastrophic forgetting, high update costs, and imprecise modification.

5.2 Memory Evolution

Memory evolution manages the repository after formation by consolidating related entries, updating conflicting knowledge, and forgetting obsolete or low-value information. These mechanisms preserve generalization, accuracy, efficiency, and timeliness, while introducing trade-offs around abstraction, overwrite decisions, and long-tail retention.

  • Memory evolution integrates new and existing memories to synthesize insights, resolve conflicts, and prune obsolete data rather than merely appending entries.
  • Consolidation merges related memories into generalized insights, updating corrects or synchronizes conflicting knowledge, and forgetting removes outdated or redundant information.
  • 5.2.1 Consolidation: Consolidation strategies range from local content merging to cluster-level fusion and global integration as memory repositories grow.
  • 5.2.1 Consolidation: Local consolidation improves precision and storage efficiency but cannot fully capture cluster-level relations or higher-order dependencies.
  • 5.2.3 Forgetting: Forgetting uses time, access frequency, and informational value to manage efficiency, but heuristic pruning can remove rarely accessed knowledge needed for correct decisions.
  • 5.2.1 Consolidation: Cluster-level fusion reorganizes related memories into higher-order reasoning units and more consistent cross-sample representations.
  • 5.2.1 Consolidation: Consolidation forms coherent long-term schemas and reduces redundancy, but information smoothing can erase outliers and unique exceptions.
  • 5.2.2 Updating: Memory updating emphasizes localized correction and synchronization, with external conflict resolution and parameter editing supporting continual self-correction.

5.3 Memory Retrieval

Memory retrieval is organized as a multi-stage pipeline covering when and why to retrieve, how to construct queries, how to search, and how to refine results. The survey also highlights increasingly autonomous retrieval decisions and the trade-off between efficient retrieval and silent failures.

  • Retrieval Timing and Intent: Automated timing lets agents decide when retrieval is necessary, progressing from fixed or instruction-driven triggers toward autonomous, self-regulated decisions.MemGPT and MemTool let the language model invoke retrieval, while ComoRAG and PRIME use failure feedback to trigger deeper retrieval.
  • Retrieval Pipeline: Memory retrieval comprises four stages: timing and intent, query construction, retrieval strategies, and post-retrieval processing.These stages determine when and why to retrieve, what to retrieve, how to search, and how to refine retrieved fragments.
  • Retrieval Timing and Intent: Automated intent selects which memory source to access, with approaches ranging from dynamic store selection to coarse-to-fine hierarchical routing.MemOS selects among parametric, activation, and plaintext memory, while H-MEM routes from domain to episode layers.
  • Retrieval Timing and Intent: Autonomous retrieval reduces unnecessary computation and noise but can cause silent failures when agents overestimate their internal knowledge and omit needed retrieval.The survey identifies a balance between supplying essential information and avoiding excessive retrieval.
  • Query Construction: Query construction improves alignment between user utterances and memory indexes through decomposition into sub-queries or semantic rewriting.Decomposition supports modular retrieval and reasoning, whereas rewriting refines query semantics before search.
  • Query Construction: Auto-RAG’s comparison of decomposition and rewriting shows that memory-retrieval query quality substantially affects reasoning performance.It evaluates HyDE and Visconde under identical retrieval conditions and selects the better strategy for each task.

6 Resources and Frameworks

The survey reviews benchmarks and frameworks for evaluating and building memory-augmented agents. It distinguishes explicit memory and lifelong-learning evaluations from broader interactive settings that impose memory demands indirectly.

  • Memory-Oriented Benchmarks: Memory benchmarks evaluate how agents construct, maintain, and exploit explicit memories across dialogues, sessions, narratives, and sometimes modalities.Table 8 organizes these benchmarks by memory focus, environment type, modality, and evaluation scale.
  • Lifelong Learning: Lifelong-learning benchmarks test continual acquisition, consolidation, and updating across long sequences of tasks or episodes.They measure preservation of useful knowledge while agents adapt to new users, domains, or interaction patterns.
  • Self-Evolving Agents: Self-evolving-agent benchmarks additionally examine self-reflection, memory editing, tool-augmented storage, and policy improvement over multiple episodes.Their focus extends beyond storing and recalling information to open-ended refinement of memory, skills, and strategies.
  • Broader Agent Benchmarks: Embodied, web, tool-use, multitask, and real-world benchmarks indirectly evaluate memory through sequential, multi-step, and long-horizon interactions.Success often requires retaining observations, goals, constraints, retrieved information, or evolving states across steps.
  • Open-Source Frameworks: Open-source memory frameworks increasingly support factual and experiential memory, multimodality, flexible structures, and reusable storage infrastructure.Table 9 compares supported memory types, architectural abstractions, multimodality, and evaluation coverage; many frameworks leave agent behavior and evaluation to applications.

7 Positions and Frontiers

The survey frames agent memory’s future as a shift from passive retrieval and hand-crafted pipelines toward generative, adaptive, self-optimizing systems. It identifies reinforcement learning, multimodal and multi-agent operation, active state maintenance, and trustworthiness as central frontiers.

  • Memory Generation: Generative memory synthesizes, compresses, and reorganizes information for current context and future utility rather than merely retrieving stored fragments.The survey emphasizes context adaptation, integration across heterogeneous signals, and learning-based optimization as desired properties.
  • Automated Memory: Manual memory strategies are interpretable and inexpensive but inflexible, often generalize poorly across dynamic environments, and tend to underperform in long-term or open-ended interactions.Existing systems commonly hand-design what to store, when to use it, and how to update or retrieve it.
  • Automated Memory: Tool-based memory management places add, update, delete, and retrieval actions inside the agent’s decision loop, supporting more coherent and transparent behavior.This approach connects memory operations directly with the agent’s reasoning process rather than separating them into external workflows.
  • Reinforcement Learning: Future RL-driven systems may invent memory organizations with less reliance on human-engineered priors and control memory formation, evolution, and retrieval end to end.The proposed direction treats memory as a learnable subsystem that coevolves with the agent through reinforcement learning.
  • Multimodal Memory: Multimodal memory must unify perception, action, environmental feedback, and other heterogeneous signals while preserving semantic alignment and temporal coherence.The survey notes that no current system provides truly omnimodal support.
  • Multi-Agent Memory: Shared memory is envisioned as an adaptive collective representation whose reads, writes, synchronization, and conflict resolution respond to agent roles, expertise, trust, and team performance.Learning-driven management is proposed for open-ended and multimodal multi-agent environments.
  • Active Memory Management: Active memory policies replace passive retention with cognitive workspaces that curate, summarize, and discard information according to task relevance.The survey presents this as a transition from data caching toward active maintenance of a coherent, queryable world state.

8 Conclusion

The conclusion presents Forms–Functions–Dynamics as a framework for clarifying agent memory beyond the long-term/short-term dichotomy. It positions memory as central to temporal coherence, continual adaptation, and long-horizon competence while identifying learnable, multimodal, multi-agent, and self-organizing systems as future directions.

  • Taxonomy: The survey organizes agent memory through forms, functions, and dynamics, replacing the coarse long-term versus short-term distinction with a broader taxonomy.Its functional categories are factual, experiential, and working memory.
  • Role of Memory: Memory supports knowledge retention, capability accumulation, and task-level reasoning, making it an essential substrate for temporal coherence, continual adaptation, and long-horizon competence.The conclusion frames memory as more than an auxiliary storage mechanism.
  • Future Directions: The survey identifies reinforcement learning, multimodal and multi-agent settings, and generative memory as directions toward fully learnable, adaptive, and self-organizing memory systems.It presents these directions as shaping the next stage of agent memory research.
  • Open Problem: Memory design remains a central and open problem in developing robust, general, and enduring artificial intelligence.The conclusion describes memory as potentially decisive for future agentic intelligence.
Loading 2512.13564v2…