PaperlayerView exact revision on arXiv

Source-linked AI summary

From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms

Jinghao Luo, Yuchen Tian, Chuxue Cao, Ziyang Luo, Hongzhan Lin, Kaixin Li, Chuyi Kong, Ruichao Yang, Jing Ma

arXiv:2605.06716v1cs.AIcs.CL

TL;DR

Research on large language model (LLM) agent memory lacks a clear technological roadmap. This survey organizes memory mechanisms into Storage, Reflection, and Experience stages. Its synthesis characterizes memory evolution as increasing information density and transforming cognitive abstraction, rather than simply expanding storage capacity.

  • Problem

    LLM agent memory research lacks a clear account of core technologies and a roadmap identifying which innovations are robust enough to build upon.

    Source: §1 · p. 2

  • Method

    The survey organizes LLM agent memory into three evolutionary stages: Storage preserves interaction trajectories, Reflection evaluates and refines records, and Experience abstracts behavioral patterns.

    Source: §1 · p. 2

  • Results

    The analysis characterizes memory evolution as increasing information density and transforming cognitive abstraction, rather than merely expanding storage capacity.

    Source: §7 · p. 9

  • Takeaways & Limitations

    Active exploration and cross-trajectory abstraction in the Experience stage support transferable behavioral experience beyond specific situations.

    Source: §7 · p. 9

    The survey lacks comprehensive quantitative performance comparisons because no unified benchmark spans all three stages and studies vary in models, environments, and prompts.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large Language Model (LLM)-based agents have fundamentally reshaped artificial intelligence by integrating external tools and planning capabilities. While memory mechanisms have emerged as the architectural cornerstone of these systems, current research remains fragmented, oscillating between operating system engineering and cognitive science. This theoretical divide prevents a unified view of technological synthesis and a coherent evolutionary perspective. To bridge this gap, this survey proposes a novel evolutionary framework for LLM agent memory mechanisms, formalizing the development process into three stages: Storage (trajectory preservation), Reflection (trajectory refinement), and Experience (trajectory abstraction). We first formally define these three stages before analyzing the three core drivers of this evolution: the necessity for long-range consistency, the challenges in dynamic environments, and the ultimate goal of continual learning. Furthermore, we specifically explore two transformative mechanisms in the frontier Experience stage: proactive exploration and cross-trajectory abstraction. By synthesizing these disparate views, this work offers robust design principles and a clear roadmap for the development of next-generation LLM agents.

1 Introduction

The survey frames LLM-agent memory evolution as three stages—Storage, Reflection, and Experience—to address fragmented paradigms and missing technological synthesis. It examines the drivers of this evolution, frontier Experience mechanisms, and future research priorities.

1 Introduction

  • Fragmentation between operating-system engineering and cognitive-science paradigms, compounded by missing technological synthesis, prevents a coherent evolutionary account and clear research roadmap.Zhang et al. (2024) emphasizes engineering-module classification, while Hu et al. (2025b) remains within static functional categories rather than explaining underlying evolutionary principles.Source: §1 · pp. 1–2
  • The proposed framework traces memory evolution through Storage for recording trajectories, Reflection for refining records, and Experience for abstracting strategies that guide future behavior.Reflection introduces dynamic evaluation loops to manage records, while Experience extracts high-level behavior patterns and strategies from clustered interactions.Source: §1 · p. 2
  • Long-range consistency, dynamic environment interaction, and continual learning are identified as the core drivers of memory evolution.LLM statelessness undermines consistency across multistep tasks and learning from prior interactions, producing recurring reasoning errors; memory mechanisms address these limitations and support self-evolution.Source: §1 · pp. 1–2
  • The survey examines proactive exploration and cross-trajectory abstraction as Experience-stage mechanisms for addressing bottlenecks in agent adaptability and autonomy.This analysis complements its examination of the structural shifts from Storage through Reflection to Experience.Source: §1 · p. 2
  • Future priorities include task-dependent dynamic memory triggering, working memory construction, comprehensive datasets especially for Experience, distributed shared memory coordination, and multimodal memory fusion.The survey identifies working memory construction as a vital core of memory mechanisms.Source: §1 · p. 2

2 Background

LLM agents select actions using instructions, observations, and retrieved external memory. The section distinguishes three evolutionary stages of memory—Storage, Reflection, and Experience—by their levels of abstraction and processing.

2.1 The LLM Agent Framework

  • An agent’s policy selects actions from instructions, current observations, and retrieved memory, distinguishing the global repository M from its context-specific retrieval m_t.The external memory repository bridges frozen parametric knowledge and evolving environmental dynamics.Source: §2.1 · p. 3

2.2 Taxonomy

  • Storage preserves trajectories with minimal transformation, maintaining a one-to-one correspondence between memory entries and execution traces.A trajectory is a chronological sequence of observation–action pairs within a task session; raw storage accumulates historical trajectories.Source: §2.2 · p. 3
  • Reflection transforms completed trajectories into independent memory entries containing critiques or corrective insights, prioritizing quality density over raw fidelity.This semantic filtering separates useful logic from trajectory-specific noise and returns processed insights to the repository for future retrieval.Source: §2.2 · p. 3
  • Experience abstracts across trajectories, compressing redundant traces into generalized schemas and extracting rules that serve as a policy prior.The inductive operator processes topologically similar trajectories and aims to satisfy the Minimum Description Length principle, supporting decisions at a higher abstraction level.Source: §2.2 · p. 3
  • Storage, Reflection, and Experience describe increasing memory abstraction rather than complete replacement: later mechanisms can retain characteristics of earlier stages.The taxonomy characterizes how memory is utilized rather than introducing additional technical categories.Source: §2.2 · p. 3

3 Evolutionary Drivers

Long-term consistency, dynamic environments, and continual learning drive memory evolution toward persistent state and goals, active knowledge management, causal reasoning, and reusable experience.

3.1 Long-Term Consistency

  • Long-term consistency drives memory mechanisms to preserve coherent states and persistent goals across extended interactions.State memory supports coherent reasoning, alignment with the external world, and stable personas. Explicit high-level goals counter local planning drift, while shared goal memory supports coordination across agents.Source: §3.1 · pp. 3–4

3.2 Dynamic Environments

  • Dynamic environments require active memory management because previously useful knowledge can become outdated while remaining semantically relevant.This motivates temporal awareness, decay policies, and more flexible retrieval rather than static storage.Source: §3.2 · p. 4
  • Delayed outcomes and cascading effects require memory to represent causal dependencies across time, not merely record interactions.Causally consistent internal worlds support robust planning.Source: §3.2 · p. 4

3.3 Continual Learning

  • Continual learning requires transforming task-specific episodic memories into reusable experience rather than merely replaying historical trajectories.Open-world deployment introduces patterns outside the training distribution; without internalizing memories into actionable knowledge, agents repeat trial and error. Experience abstraction provides guidance across future scenarios.Source: §3.3 · pp. 4–5
  • Long-running agents need strategic memory addition and deletion rather than unrestricted storage expansion.Extended interaction produces linear storage growth (Hu et al., 2023; Packer et al., 2023). Unrestricted expansion can impair performance as errors propagate through memory and undermine learning (Srivastava and He, 2025).Source: §3.3 · p. 5

4 Evolutionary Path

The survey frames agent memory evolution as faithful trajectory preservation in Storage, correction and denoising in Reflection, and high-order abstraction in Experience.

4.1 Storage

  • Storage preserves interaction trajectories to maintain behavioral consistency despite limited context windows, using linear, vector, or structured representations.Linear storage uses chronological tokens and FIFO, with context extension and mechanical noise reduction; vector storage emphasizes semantic retrieval or weighting by temporal decay and importance. Structured storage uses relational tables, memory hierarchies, and entity–relation networks to address capacity limits, retrieval-speed trade-offs, and retrieval ambiguity.Source: §4.1 · p. 5

4.2 Reflection

  • Reflection improves memory quality by correcting and denoising trajectories through introspection, environmental feedback, and collaboration.Introspection uses internal model knowledge without external feedback for error correction, memory lifecycle management, and trajectory compression. Environmental outcomes anchor reflection against hallucinations while refining behavioral policies and world models; heterogeneous agents use role division and consensus to overcome individual cognitive limits.Source: §4.2 · pp. 5–6

4.3 Experience

  • Experience abstracts context-dependent trajectories into reusable heuristics, compressing memory and enabling generalization beyond the fragmented memories produced by reflection.Reflected memories can impose substantial retrieval and inference costs, while imitation of successful trajectories can fail under minor contextual changes. The survey’s abstraction-based definition excludes some research that uses the term experience.Source: §4.3 · p. 6
  • Explicit experience converts trajectory clusters into human-readable, editable abstractions, including natural-language policies, executable entities, and evolving skill libraries.Skill libraries couple procedural abstraction with induction, reuse, and refinement (Zhang et al., 2026; Ni et al., 2026).Source: §4.3 · p. 6
  • Implicit experience internalizes interaction histories through model parameters or hidden-layer latent variables, addressing explicit memory’s inference overhead and context limitations.Fine-tuning converts experience into intrinsic model capabilities; latent representations instead support dynamic invocation during inference.Source: §4.3 · p. 6
  • Hybrid experience cycles between explicit accumulation and implicit internalization, treating explicit experience as a high-capacity cache periodically compressed into model weights.Experience transfer occurs through periodic parameter updates.Source: §4.3 · p. 6

5 Transformative Experience

Experience combines goal-driven active exploration with cross-trajectory abstraction to acquire experience and derive policy priors for unseen scenarios. Their feedback loop supports autonomous, continuous agent evolution.

5.1 Active Exploration

  • Active exploration turns passive information recorders into goal-driven experience collectors, with prior experience guiding exploration and exploration outcomes abstracted back into memory.Exploration drivers extend beyond randomness to reward signals targeting valuable states, curricula of increasing task difficulty, and reuse of historical trajectories.Source: §5.1 · pp. 6–7
  • Experience extracts a separate rule set serving as a policy prior for unseen scenarios, whereas Reflection refines stored trajectories to assist similar future tasks.This marks a shift from trajectory-local refinement to cross-trajectory abstraction.Source: §5.1 · p. 7, Table 1 · p. 7
  • Exploration expands capabilities across breadth, depth, and strategy: addressing unfamiliar environments, developing higher-order task skills, and improving long-horizon decision paths.Breadth uses curiosity to structure experience, depth develops complex experiential strategies, and strategic exploration uses accumulated experience to improve planning precision.Source: §5.1 · p. 7

5.2 Cross-Trajectory Abstraction

  • Cross-trajectory abstraction derives generalizable experience through success–failure comparison, behavioral sequence aggregation, reusable program functions, and fine-tuning model parameters.Unlike reflection’s correction of errors within individual trajectories, these mechanisms perform induction across trajectories to extract patterns that guide unknown tasks.Source: §5.2 · pp. 7–8
  • Abstraction progresses from natural-language rules through modular execution skeletons to model weights, determining experience’s generalization boundaries and interpretability.Shallow abstraction preserves semantic logic; intermediate abstraction removes natural-language redundancy; deep abstraction compresses trajectory distributions into model weights for intuitive decision-making.Source: §5.2 · p. 8
  • Exploration and abstraction form the central feedback loop driving autonomous, continuous agent evolution during the Experience stage.Highly abstract memories can also help agents avoid challenges associated with conventional memory mechanisms (Gharat et al., 2025; Srivastava and He, 2025).Source: §5.2 · p. 8

6 Future Directions

Future directions include on-demand memory access, better working-memory organization, and benchmarks for experience abstraction and generalization. The section also calls for consensus-based shared memory and multimodal memories with unified temporal and semantic structure.

Active Memory Perception

  • Memory mechanisms should autonomously decide whether additional memory is needed and which type to retrieve, rather than relying on passive triggering.Irrelevant or obsolete memories can disrupt reasoning (Tan et al., 2025). Autonomous retrieval controllers have begun addressing this challenge.Source: §6 · p. 8

Organization of Working Memory

  • Working-memory research should investigate interval isolation, critical-decision-node integration, and adaptive pruning (Sun et al., 2025c; Zhang et al., 2025i; Nan et al., 2025).Increasing task complexity and longer horizons make working-memory construction a bottleneck. Reconstructing past trajectories into dynamic, plastic memory intervals supports attention allocation (Hu et al., 2024; Luo et al., 2025a).Source: §6 · p. 8

Benchmark for Experience

  • Experience benchmarks should evaluate abstraction and generalization, which existing retrieval- and denoising-focused datasets inadequately assess.Evaluating the experience lifecycle is linked to meta-learning and self-evolution through active generalization (Behrouz et al., 2024). The survey’s evolutionary framework offers a foundation for developing these benchmarks.Source: §6 · p. 8

Distributed Shared Memory

  • Distributed shared memory for multi-agent collaboration should develop consensus systems that synchronize individual perspectives with collective knowledge.Explicit dialogue creates bandwidth bottlenecks and exchange noise (Tran et al., 2025; Liao et al., 2025). Consensus systems aim to support more agile socialized experience evolution (Rezazadeh et al., 2025).Source: §6 · pp. 8–9

Multimodal Memory

  • Multimodal memory research should pursue multimodal abstraction, cross-modal temporal alignment, and efficient consolidation.Visual, linguistic, and other perceptual states require memory units with unified temporality and semantics (Liu et al., 2025d; Zhou et al., 2025a; He et al., 2025). For embodied intelligence, internal world-model integrity directly influences planning and execution (Feng et al., 2025b; Long et al., 2025b).Source: §6 · p. 9

7 Conclusion

The survey frames LLM-agent memory evolution as progression through storage, reflection, and experience, driven by richer information and changing cognitive abstraction rather than storage expansion alone.

Conclusion

  • Memory mechanisms evolve through three progressive stages: storage, reflection, and experience.The survey offers this framework to guide the design of more advanced LLM-agent memory mechanisms.Source: §7 · p. 9
  • Memory evolution fundamentally enhances information density and transforms cognitive abstraction, rather than merely expanding storage capacity.Source: §7 · p. 9
  • Active exploration and cross-trajectory abstraction enable experience-stage agents to transcend situational constraints and acquire transferable behavioral experience.Source: §7 · p. 9

Limitations

The survey lacks comprehensive quantitative comparisons, and its experience taxonomy intersects with established learning paradigms rather than defining an entirely new one. Its coverage also risks underrepresenting early contributions and includes recent preprints without formal peer review.

Lack of Direct Quantitative Comparison

  • The survey provides qualitative analysis rather than comprehensive performance comparisons because stage objectives differ, unified benchmarks are absent, and study configurations vary.No unified benchmark evaluates storage, reflection, and experience comprehensively. Differences in foundation models, environments, and prompts make direct numerical comparisons potentially misleading.Source: Limitations · p. 9

Relation to Established Learning Paradigms

  • Experience, particularly implicit experience, intersects with fine-tuning, reinforcement learning, and meta-learning rather than constituting an entirely novel learning paradigm.The taxonomy emphasizes these techniques’ deployment within memory-centric LLM agent architectures, where experience intermediates between interaction trajectories and parameter updates.Source: Limitations · p. 9

Temporal Coverage and Recency Bias

  • Rapid research growth creates recency bias: early influential works may receive insufficient attention, while some included recent preprints lack formal peer review.Memory research grew rapidly during 2024–2025, with experience emerging as a coherent direction only in the latter half of 2025. The survey prioritizes novel architectures or reproducible results to balance rigor and timeliness.Source: Limitations · p. 9

A Overview · B Detail within the Evolutionary Path

The survey organizes LLM agent memory into Storage, Reflection, and Experience according to the depth of historical-trajectory utilization. Advancing toward Experience requires active exploration and cross-trajectory abstraction.

A Overview

  • The framework classifies memory by historical-trajectory utilization and organizes analysis of evolutionary drivers, development pathways, and frontier characteristics.It groups the drivers of changes in memory mechanisms and agents’ capabilities into three dimensions.Source: A Overview · pp. 20–21
  • Storage faithfully preserves long-term interaction trajectories to address LLM agents’ memory-capacity constraints.It forms the foundational stage of the evolutionary framework.Source: A Overview · p. 21
  • Reflection introduces dynamic evaluation loops, turning memory from an information recorder into an evaluator that mitigates hallucinations and logical errors.Source: A Overview · p. 21
  • Experience abstracts higher-order behavioral patterns across multiple trajectories, compressing redundant memory into transferable, reusable heuristic strategies.The survey positions Experience as the highest cognitive stage in its framework.Source: A Overview · p. 21
  • Active exploration and cross-trajectory abstraction shift agents from passive information recipients to goal-driven experience collectors, supporting proactive generalization to unknown tasks.The survey identifies these mechanisms as pivotal technological shifts required to advance toward Experience.Source: A Overview · p. 21
  • The survey combines overlooked research, frontier work through 2025, and broader theoretical perspectives while identifying underexplored or unresolved issues.It presents itself as the first systematic evolutionary survey of memory mechanisms, aiming to bridge fragmented research and provide a foundation for further exploration.Source: A Overview · p. 21

B Detail within the Evolutionary Path

  • The detailed evolutionary-path section explains representative works at each stage through their core contributions, technical mechanisms, and positions within the evolutionary trajectory.Source: B Detail within the Evolutionary Path · p. 21

B.1 Storage · B.2 Reflection

Storage preserves interaction trajectories through linear, vector, and structured architectures but does not ensure memory quality. Reflection refines those trajectories through introspection, environmental feedback, and multi-agent coordination.

B.1 Storage

  • Storage prioritizes faithful trajectory preservation through linear, vector, and structured architectures, but remains susceptible to model stochasticity and hallucinations.Preserving trajectories supports accurate perceptions of internal and external states, while supplying context for continuity and reasoning.Source: §B.1 · pp. 21–22
  • Linear storage maintains chronological token streams with FIFO management, extending usable context through context-window adaptation and information sparsification.Context-window adaptation modifies attention, positional encoding, or input structures without changing trajectory semantics. Sparsification mechanically removes low-utility tokens using statistical or attention-based heuristics, improving efficiency without reflection or semantic abstraction.Source: §B.1 · p. 22
  • Vector storage encodes trajectories in high-dimensional spaces, shifting the central challenge from capacity toward retrieving task-relevant memories.Semantic retrieval uses embedding proximity but treats recalled content as flat historical evidence. Weighted retrieval adds signals such as temporal decay, recency, and importance, improving prioritization without introducing abstraction.Source: §B.1 · pp. 22–23
  • Structured storage organizes relationships through tabular databases, tiered architectures, and semantic graphs to support precise operations, logical reasoning, and multi-hop retrieval.Databases provide relational storage and SQL access, while tiered architectures balance capacity and access latency. Semantic graphs encode entities and relations for precise updates, logical constraints, environment-centric reasoning, and efficient traversal.Source: §B.1 · p. 23

B.2 Reflection

  • Reflection addresses accumulated errors in stored trajectories by using introspection, environmental feedback, and coordination to rectify and denoise memory.Raw interaction histories mix successful sequences with hallucinations, logical errors, and invalid attempts; preserving them without evaluation can perpetuate failures.Source: §B.2 · p. 23
  • Introspective reflection uses the agent’s own knowledge for error rectification, dynamic memory maintenance, and knowledge compression without external feedback.Self-critique verifies and repairs trajectories, while lifecycle maintenance updates knowledge schemas and entity relations. Compression distills trajectories into compact, reusable representations, including procedural memories and task-aligned abstractions.Source: §B.2 · pp. 23–24
  • Environmental reflection calibrates memory against real-world outcomes through environment modeling and decision optimization, countering introspection’s potential inconsistency with factual reality.Environment modeling updates knowledge about world rules, tools, and preferences as external conditions change. Decision optimization learns memory-management policies from rewards or execution feedback and validates or prunes planning hierarchies.Source: §B.2 · p. 24
  • Collaborative reflection combines specialized agent roles and consensus to reduce individual cognitive bottlenecks and hallucinations when processing complex memory trajectories.Multi-dimensional calibration distributes memory management across heterogeneous agents, combining memory types and specialized processing roles. Cross-agent integration and personalized feedback support narrative coherence, reasoning consistency, and collaboration efficiency.Source: §B.2 · p. 24

B.3 Experience · C Extended Discussion on Multimodal Memory Mechanisms · C.1 Current Approaches.

Experience memory abstracts transferable knowledge from trajectories through explicit, implicit, and hybrid mechanisms. Multimodal memory remains predominantly storage-focused, with representation and retrieval approaches balancing semantic abstraction against fidelity to original modalities.

B.3 Experience

  • Experience abstracts context-independent skills or rules from trajectory clusters, moving beyond reflection’s trajectory-level corrections toward transferable strategies.Reflection mitigates hallucinations and noise but can expand memory banks linearly, increasing inference burdens without producing transferable strategic knowledge.Source: §B.3 · pp. 24–25
  • Explicit experience distills trajectories into readable heuristic guidelines or executable procedural primitives, supporting interpretable reuse and reducing inference or planning overhead.Guidelines encompass textual rules, experience graphs capturing logical dependencies, and compositional knowledge modules. Procedural primitives encapsulate frequent action sequences as callable functions that can be composed into complex behaviors.Source: §B.3 · p. 25
  • Implicit experience encodes interaction histories as latent variables or model parameters; latent modulation incorporates historical insights into reasoning without updating model weights.MemGen dynamically injects latent token sequences conditioned on the current reasoning state (Zhang et al., 2025c). Another approach alternates Fast Retrieval and Slow Integration within latent space (Zhang et al., 2025d).Source: §B.3 · pp. 25–26
  • Parameter internalization uses gradient updates to transform explicit trajectories into model capabilities through context distillation and reinforcement learning.Distillation internalizes corrective hints and business rules, while Experience Stripping removes retrieval segments during training to encourage autonomous reasoning (Alakuijala et al., 2025; Liu et al., 2025b). Reinforcement-learning approaches internalize trial-and-error experience, reflections, multi-turn interactions, and tool-use strategies.Source: §B.3 · p. 26
  • Hybrid experience cycles between explicit accumulation and parameter internalization to address repository growth, retrieval latency, and updates that lag environmental changes.Experience Transfer progressively reduces retrieval dependence by distilling structured experience into policy parameters. Replay-buffer approaches retain valuable exploration trajectories for immediate use while on-policy and off-policy updates internalize successful experience.Source: §B.3 · p. 26

C Extended Discussion on Multimodal Memory Mechanisms

  • Multimodal memory encompasses text, images, audio, and audiovisual video to capture cross-modal dependencies in environments such as embodied navigation and interactive web browsing.These environments require agents to combine linguistic reasoning with perception across multiple modalities.Source: C Extended Discussion on Multimodal… · p. 26

C.1 Current Approaches.

  • Multimodal representation is shifting from predominant textualization toward hybrid representations that preserve original-modality information alongside semantic summaries.Textualization loses information that cannot be adequately expressed in language. Dense original-modality representations stored alongside text, multi-path indexing, adaptive retrieval, and shared-space embeddings help negotiate semantic utility versus perceptual fidelity.Source: §C.1 · p. 27
  • Multimodal retrieval uses parallel fusion, hierarchical search, or agent-directed queries to balance semantic abstraction and perceptual fidelity.Research remains predominantly at Storage, with limited Reflection and Experience work. Parallel fusion combines semantic, lexical, and visual rankings; hierarchical search moves from summaries to original modalities; agent-driven retrieval iterates modality-specific queries until sufficient information is obtained.Source: §C.1 · pp. 26–27

C.2 Unique Challenges.

Multimodal memory faces distinctive challenges in aligning heterogeneous signals, maintaining temporal consistency, and consolidating or forgetting information. Research remains concentrated in Storage, with Reflection and Experience comparatively underexplored.

Multimodal Alignment

  • Multimodal alignment must bind visual, auditory, and textual signals into unified semantic units despite differences in granularity, semantic density, and representation.These cross-modal asymmetries complicate alignment (Cai et al., 2025a; Wang et al., 2025a). Misalignment can prevent complete recall and introduce irrelevant signals that impede reasoning.Source: §C.2 · p. 27

Temporal Consistency

  • Temporal consistency requires adaptive segmentation, multiscale indexing, and hierarchical consolidation because multimodal events unfold continuously and their boundaries need not match physical time.Continuous-time memory makes unit length and granularity important (Lian et al., 2026). A single action can span multiple visual frames and extended speech segments.Source: §C.2 · p. 27

Consolidation and Forgetting

  • Multimodal consolidation and forgetting must accommodate incompatible similarity measures, modality-dependent validity periods, and potential loss of nontextual perceptual detail.Cross-modal similarity lacks a unified metric; persistent textual preferences versus rapidly outdated scene states complicate decay functions. Consolidation can also discard nontextual perceptual details (Lian et al., 2026).Source: §C.2 · pp. 27–28

C.2 Unique Challenges.

  • Multimodal memory research remains predominantly at Storage, with scarce Reflection and Experience work indicating an early developmental stage.Future directions include addressing cross-modal errors and hallucinations through Reflection and extracting modality-invariant policy priors from multimodal trajectory clusters through Experience (Allard et al., 2026).Source: §C.2 · p. 28

D Datasets and Benchmarks

The survey organizes memory benchmarks into Storage, Reflection, and Experience stages. Existing evaluations still primarily assess static storage and retrieval, leaving other memory capabilities in dynamic interaction insufficiently evaluated.

Storage Stage

  • Extreme-context benchmarks probe effective memory limits by testing retrieval and processing of small facts amid massive distracting contexts, including visual information.These tests belong to the Storage stage, which evaluates accurate long-distance storage and retrieval across scenarios, tasks, and modalities.Source: D Datasets and Benchmarks · p. 28
  • Interactive-consistency benchmarks assess whether agents maintain coherent memory across sessions, including knowledge updates, rejection capabilities, and long-term persona consistency.This category emphasizes memory consistency over long interaction histories rather than isolated retrieval.Source: D Datasets and Benchmarks · p. 28
  • Relational-fact benchmarks assess semantic association, multihop reasoning, cross-document fact integration, and memory recall in proactive and passive interaction.Their scope includes personal trivia, factual evidence integration, and emotional-support scenarios.Source: D Datasets and Benchmarks · p. 28

Reflection Stage

  • Error-correction benchmarks evaluate memory fidelity by detecting errors and hallucinations throughout the memory lifecycle.Reflection evaluates how raw trajectories become high-quality memory, emphasizing denoising, user alignment, and complex-environment perception. Error-correction tests examine memory search, editing, matching, extraction, and updating.Source: D Datasets and Benchmarks · pp. 28–29
  • Personalization benchmarks assess reflection-based extraction of implicit preferences, user traits, key information, and shared components from interaction histories.The emphasis is on discovering latent personalization information rather than merely recalling explicitly stored facts.Source: D Datasets and Benchmarks · p. 29
  • Dynamic-reasoning benchmarks assess memory-supported multistep reasoning and perception in complex environments.Coverage includes selective forgetting, decision backtracking, real-world scenarios, and summarization and transition mechanisms.Source: D Datasets and Benchmarks · p. 29

Experience Stage

  • Experience-stage benchmarks are scarce and assess whether agents abstract, internalize, and transfer generalizable experience from fragmented trajectories in dynamic environments.Wu et al. (2024) and Ai et al. (2025) evaluate experience extraction and internalization through deployment-like feedback cycles. Wei et al. (2025a) and Zheng et al. (2025) assess abstraction and generalization through cross-task transfer.Source: D Datasets and Benchmarks · p. 29

Related papers

Loading 2605.06716v1…