Memory in the Age of AI Agents

• Agent memory enables otherwise static LLMs to adapt continually through environmental interaction and supports applications requiring historical-information management, including personalized chatbots, recommender systems, social simulations, and financial investigations.The survey grounds continual evolution in agent memory and identifies memory management as necessary beyond ephemeral behavior.Source: §1 · pp. 4–5
• Existing taxonomies no longer capture rapid advances or the fragmented implementations, objectives, assumptions, and terminology now grouped under “agent memory.”Underrepresented examples include reusable tools distilled from past experiences and memory-augmented test-time scaling methods.Source: §1 · p. 5
• The survey establishes a systematic framework that reconciles definitions, connects emerging trends, and examines agent memory through forms, functions, and dynamics.Its questions cover definitions and relationships to LLM memory, RAG, and context engineering; representational forms; functional purposes; temporal operation; and future frontiers.Source: §1 · p. 5
• The taxonomy distinguishes token-level, parametric, and latent memory forms; factual, experiential, and working memory functions; and memory formulation, retrieval, and evolution dynamics.Factual memory records interaction knowledge, experiential memory improves problem solving through task execution, and working memory manages workspace information during individual tasks.Source: §1 · p. 5
• The survey further compiles benchmarks and open-source frameworks and examines emerging frontiers including memory automation, reinforcement learning integration, multimodal memory, multi-agent shared memory, and trustworthiness.These resources and directions are intended to support knowledge sharing, future development, and further exploration by researchers and practitioners.Source: §1 · pp. 5–6
• Its contributions include analyzing how memory forms and functional purposes align with distinct agentic objectives while identifying future opportunities for advancing agent memory research.The survey presents this alignment analysis alongside its multidimensional taxonomy and resource collection.Source: §1 · pp. 5–6

2 Preliminaries: Formalizing Agents and Memory

• The framework covers both single-agent and multi-agent systems, modeling trajectories as repeated observation, optional memory retrieval, LLM computation, and action execution.It unifies within-task short-term and across-task long-term memory through read/write interactions in the agent loop.Source: §2 · p. 6, §2.1 · p. 7

2.1 LLM-based Agent Systems

• LLM agents act through heterogeneous spaces that include language generation, tool invocation, planning, environment control, and communication, all produced by an autoregressive LLM policy.Internal reasoning, latent deliberation, and scratchpad computation are abstracted away from the executable action.Source: §2.1 · pp. 6–7

2.2 Agent Memory Systems

• Agent memory is modeled as an evolving state that supplements insufficient instantaneous observations and may simultaneously store cross-task information and task-specific short-term content.The memory state can be a text buffer, key–value store, vector database, graph, or hybrid representation, without requiring separate short- and long-term modules.Source: §2.2 · p. 7
• Memory operates through formation, evolution, and retrieval: artifacts are selectively converted into candidates, integrated by consolidation or restructuring, and returned as task-aware signals for action selection.Formation avoids storing complete histories, while evolution can resolve conflicts, discard low-utility information, and improve retrieval efficiency.Source: §2.2 · pp. 7–8
• Short-term and long-term memory arise from the temporal schedules of formation, evolution, and retrieval rather than from discrete architectural modules.Operations may occur minimally, intermittently, continuously, episodically at task boundaries, or through iterative refinement, producing systems from passive buffers to evolving knowledge bases.Source: §2.2 · p. 8

2.3 Comparing Agent Memory with Other Key Concepts

• Agent memory overlaps with LLM memory, RAG, and context engineering but differs in scope, temporal characteristics, and functional role, motivating explicit conceptual separation.Figure 2 summarizes shared implementations while characterizing agent memory as persistent and self-evolving, unlike architectural optimization, static knowledge access, or transient resource management.Source: §2.3 · p. 8, Figure 2 · p. 9

2.3.1 Agent Memory vs. LLM Memory

• Most historical “LLM memory” work can be reinterpreted as agent memory, whereas mechanisms that modify internal model state, such as KV-cache management or recurrent-state persistence, remain outside its operational scope.The distinction is whether the method equips an agent with evolving, cross-task memory operations rather than expanding the model’s intrinsic representational capacity.Source: §2.3.1 · pp. 9–10

2.3.2 Agent Memory vs. RAG

• RAG and context engineering share substantial mechanisms with agent memory, but practical distinctions center on persistence and purpose: RAG usually serves task-specific external knowledge access, while context engineering optimizes finite-window resources.Agent memory maintains an internal, persistent, self-evolving base across interaction, although dynamic RAG, graph systems, and working-memory techniques increasingly blur the boundaries.Source: §2.3.2 · pp. 10–11

2.3.3 Agent Memory vs. Context Engineering

• Context engineering organizes the interaction interface between LLMs and operational environments, including tool-integrated reasoning, selection pipelines, and communication protocols.It focuses on correctly formatting and scheduling instructions, tool calls, and intermediate states within context-window constraints.Source: §2.3.3 · p. 12
• Context engineering emphasizes resource allocation, syntactic validity, and execution efficiency rather than persistent changes to the agent’s knowledge or state.Source: §2.3.3 · p. 12
• Agent memory persistently stores factual knowledge, accumulates experiential traces (Zhao et al., 2024; Tang et al., 2025d; Zhang et al., 2025d), and can internalize memory into model parameters (Wang et al., 2025o).Source: §2.3.3 · p. 12
• Agent memory governs what the agent knows, has experienced, and how these elements evolve over time, beyond transient context assembly.Source: §2.3.3 · p. 12
• Memory systems can consolidate repeated interactions into knowledge (Tan et al., 2025c) and abstract procedural knowledge from past successes and failures (Ouyang et al., 2025).Source: §2.3.3 · p. 12
• Agent memory maintains a coherent identity across tasks and episodes (Wang et al., 2024f), forming an internal substrate for learning, adaptation, autonomy, and persistent cognition beyond one context window.Source: §2.3.3 · p. 12

3 Form: What Carries Memory?

• Different task settings call for different memory storage forms, whose structural properties shape how agents accumulate information, maintain behavioral consistency, and operate across varied scenarios.Source: §3 · p. 12
• The section classifies agent memory into three forms—token-level, parametric, and latent—based on where memory resides and how it is represented.These forms establish the structural framework for understanding what carries memory.Source: §3 · p. 13

3.1 Token-level Memory

• Token-level memory uses explicit, discrete units that remain externally visible and can be individually accessed, modified, reconstructed, and stored structurally over time.Source: §3 · p. 13

3.2 Parametric Memory

• Parametric memory stores information in model parameters, encoding it through statistical patterns accessed implicitly during forward computation.Source: §3 · p. 13

3.3 Latent Memory

• Latent memory represents information in hidden states, continuous representations, or evolving latent structures that persist and update during or across interaction cycles.It captures context-dependent internal states.Source: §3 · p. 13

3.1 Token-level Memory

• Token-level memory stores discrete, externally accessible units—including text, visual, and audio elements—and is organized as flat, planar, or hierarchical topologies.Its explicit representation supports inspection, editing, retrieval, routing, conflict handling, and coordination with parametric or latent memory.Source: §3.1 · pp. 13–14, Figure 3 · p. 14
• Across token-level memory, increasing topological complexity governs how agents search, update, reason over, and structure stored information.The survey presents representative methods comparatively in Table 1 across modality support, functional type, memory content or structure, and application task.Source: §3.1 · p. 14, Table 1 · pp. 15–17

3.1.1 Flat Memory (1D)

• Flat Memory accumulates independent units without explicit relationships, enabling scalable broad recall and lightweight updates but limiting coherence, compositional reasoning, planning, and abstraction.Similarity retrieval provides flexible access, while growing collections can accumulate redundancy, noise, and contextually disconnected information.Source: §3.1.1 · pp. 15, 20
• Flat-memory research spans dialogue, preference, profile, experience, and multimodal storage, increasingly compressing, structuring, reflecting on, and adaptively optimizing accumulated information.Experiential systems range from archived trajectories to generalized workflows, executable skills, and dynamically maintained knowledge; multimodal systems store discrete units from images, video, audio, and text.Source: §3.1.1 · p. 19

3.1.2 Planar Memory (2D)

• Planar Memory encodes associations among units within one graph, tree, table, or related layer, enabling structured lookup and relational traversal beyond simple iteration.Its monolithic, non-hierarchical organization incurs increasing redundancy, construction cost, and search cost as task complexity grows.Source: §3.1.2 · pp. 20–21

3.1.3 Hierarchical Memory (3D)

• Hierarchical Memory connects multiple abstraction layers, allowing lateral and vertical navigation from raw observations through summaries to higher-level patterns.Pyramid designs support coarse-to-fine retrieval, while multi-layer designs specialize modules or levels by information type or function.Source: §3.1.3 · p. 21
• Hierarchical token-level memory combines raw dialogue, compressed summaries, relational graphs, and information-specific layers to support structured, multi-hop retrieval.Examples include HippoRAG, HippoRAG 2 (Gutiérrez et al., 2025), AriGraph (Anokhin et al., 2024), SGMem (Wu et al., 2025h), and CAM (Li et al., 2025g). CompassMem (Hu et al., 2026b) and MAGMA (Jiang et al., 2026) further explore logical, hierarchical composition.Source: §3.1.3 · p. 22
• By intersecting hierarchical and relational dimensions, this memory supports complex cross-layer queries and holistic contextualization, but dense organization complicates efficient retrieval and semantic preservation.Designing an optimal three-dimensional layout and ensuring that stored memories remain semantically meaningful are identified as critical unresolved problems.Source: §3.1.3 · p. 22

3.2 Parametric Memory

• Internal and external parametric memory are distinguished by whether information is absorbed into the base model or attached modularly alongside it.Internal memory has simple structure and no extra inference or deployment cost but requires costly retraining and risks forgetting, making it better suited to domain knowledge or task priors than short personalized or working memory.Source: §3.2 · p. 22, §3.2.1 · p. 24

3.2.1 Internal Parametric Memory

• Internal parametric memory directly modifies the model’s original weights or biases, with injection occurring during pre-training, mid-training, or post-training.Representative methods store retrieval mechanisms, agent experience, personalized histories, task outcomes, or edited knowledge in internal parameters; this adds no extra parameters or modules.Source: §3.2.1 · pp. 23–24

3.2.2 External Parametric Memory

• External parametric memory stores knowledge in auxiliary modules or models, enabling modular updates, personalization, replacement, and rollback without altering the base model.Adapters, LoRA modules, routing mechanisms, and separate memory models support reversible storage while reducing catastrophic interference, although effectiveness depends on integration with the model’s internal representation flow.Source: §3.2.2 · pp. 24–25

3.3 Latent Memory

• Latent memory resides in internal representations such as KV caches, activations, hidden states, and latent embeddings rather than human-readable tokens or dedicated parameter sets.This form avoids exposing memory in plaintext and may reduce inference latency while preserving fine-grained contextual signals.Source: §3.3 · p. 26

3.3.1 Generate

• Generate methods create compact latent representations from contexts, trajectories, or multimodal inputs, then store or condition later inference on them as reusable memory units.They extend usable context, preserve task-specific intermediate states across episodes, and reduce repeated full-context processing, but generation can cause information loss, drift, bias, and added training complexity.Source: §3.3.1 · pp. 26, 28

3.3.2 Reuse

• Reuse methods directly carry forward internal activations, especially KV pairs, using retrieval-aware organization to support long-range access without modifying stored states.They preserve activation fidelity and integrate simply, but raw caches grow with context length, making memory use and retrieval efficiency dependent on indexing strategies.Source: §3.3.2 · p. 28

3.3.3 Transform

• Transform methods modify, compress, or restructure existing latent states through selection, aggregation, retrieval enhancement, or re-encoding instead of generating new states or replaying raw KV caches.These transformations produce compact, information-dense memories that extend usable context and improve reasoning, while introducing information loss, reduced interpretability, and additional computation.Source: §3.3.3 · p. 28, Latent Memory · p. 29

Adaptation

• Choosing a memory type encodes desired agent behavior and should therefore be guided by task characteristics rather than treated as a simple combinatorial choice.The section presents idealized task–memory matches as guidance, while noting that each scenario may benefit from other memory types.Source: §3.4 · p. 30

Token-level Memory

• Token-level memory supports explicit, controllable, auditable, and stable knowledge management, especially for high-frequency updates, structured information, and high-stakes domains.Its externalizability enables inspection, transfer, and precise add/delete/update operations, while plug-and-play integration avoids modifying foundation-model parameters.Source: §3.4 · p. 30, Token-level Memory · p. 30

Parametric Memory

• Parametric memory internalizes general knowledge and rules, making it suitable for abstraction, fluid reasoning, generalization, and consistent behavioral or stylistic patterns.Representative applications include role-playing or persona-consistent behaviors (Shao et al., 2023; Zhou et al., 2024a), mathematical reasoning, coding, games, alignment, and expert responses.Source: Token-level Memory · p. 30, Parametric Memory · p. 30

Latent Memory

• Latent memory prioritizes privacy, semantic capacity, retrieval efficiency, and scalability, making it suited to multimodal, constrained, dynamic, and privacy-sensitive environments.Its low readability protects sensitive information, while compact representations support large-scale knowledge injection and cross-modal or cross-task correlations; examples include multimodal architectures (Shi et al., 2025a; Cheng and Schwing, 2022; Wu et al., 2025d).Source: Parametric Memory · p. 31, Latent Memory · p. 31

4 Functions: Why Agents Need Memory?

• The taxonomy distinguishes factual memory for explicit knowledge, experiential memory for improvement from past trajectories, and working memory for active context management.These functions span long-term persistence and short-term reasoning, connected through encoding, processing, and retrieval.Source: §4 · pp. 31–32

4.1 Factual Memory

• Factual memory transforms explicit facts about users, events, and environments into reusable knowledge that supports consistency, coherence, and adaptability.Systems process episodic traces through summarization, reflection, extraction, and induction, storing abstractions in vector databases, key-value stores, or knowledge graphs.Source: §4.1 · pp. 32–33

4.1.1 User factual memory

• User factual memory maintains identities, preferences, goals, constraints, and commitments across sessions, reducing repetition, intent drift, and contradictory responses.Retrieval ranking, semantic abstraction, and explicit task-state tracking improve long-horizon coherence, task success, and operational efficiency.Source: §4.1.1 · pp. 35–36

4.1.2 Environment factual memory

• Environment factual memory provides an updatable, retrievable, and governable external fact layer for long-horizon reasoning and multi-agent coordination.Its knowledge-persistence axis supports completeness, interpretability, and editability, while shared access maintains cross-agent and cross-stage consistency.Source: §4.1.2 · p. 36

4.2 Experiential Memory

• Experiential memory stores trajectories, distilled strategies, and interaction outcomes to support continual learning and self-evolution across episodes.By converting feedback into reusable knowledge, it enables non-parametric adaptation, error correction, heuristic abstraction, and routine-skill compilation without frequent parameter updates.Source: §4.2 · pp. 37–38
• Experiential memory spans raw cases, abstract strategies, executable skills, and hybrid systems that combine these levels to balance evidence, generalization, and action.Cases preserve experience, strategies guide planning, skills operationalize execution, and hybrid architectures dynamically connect these components.Source: Table 5 · p. 39, §4.2.2 · p. 41, §4.2.3 · p. 42

4.2.1 Case-based Memory

• Case-based memory preserves minimally processed trajectories or solutions for replay and verifiable imitation, but incurs retrieval and context-window costs without providing orchestration logic.Examples include trajectory replay, successful solution exemplars, code playbooks, and episodic records of actions and outcomes.Source: §4.2.1 · pp. 39–40

4.2.2 Strategy-based Memory

• Strategy-based memory distills experiences into insights, workflows, and reasoning patterns that constrain generative search and improve transfer, while remaining non-executable planning guidance.Strategies provide abstract planning logic; callable capabilities and tools are handled by skill-based memory.Source: §4.2.2 · pp. 40–41

4.2.3 Skill-based Memory

• Skill-based memory forms the agent’s active execution substrate by converting strategies into callable, verifiable, composable procedures spanning code, scripts, APIs, MCP interfaces, and learnable architectures.Its lifecycle includes capability synthesis, retrieval, refinement, interoperability, and execution.Source: §4.2.3 · pp. 41–42

4.2.4 Hybrid memory

• Hybrid memory integrates raw episodes, distilled rules, and executable skills, enabling systems to retrieve concrete evidence, apply general heuristics, and progressively compile repeated successes into efficient capabilities.Examples include hierarchical case–strategy retrieval, self-updating case-and-skill libraries, and evolutionary transitions from retrieval to execution.Source: §4.2.4 · p. 42

4.3 Working Memory

• Working memory transforms the context window from a passive buffer into a controllable, writable, interference-resistant workspace that actively selects, maintains, and rewrites task-relevant information within an episode.This increases task-relevant information density, suppresses redundancy, and supports coherent reasoning under fixed attention budgets.Source: §4.3 · pp. 42–43

4.3.1 Single-turn Working Memory

• Single-turn working memory compresses massive immediate inputs through token condensation and observation abstraction, increasing information density while addressing static inputs rather than temporal interaction continuity.Hard, soft, and hybrid condensation trade token efficiency against dependency preservation and detail retention; abstraction converts observations into fixed-size semantic states.Source: §4.3.1 · pp. 43, 45

4.3.2 Multi-turn Working Memory

• Multi-turn working memory maintains operable state over long interactions through state consolidation, hierarchical folding, and cognitive planning, preserving temporal coherence and goal alignment under bounded resources.These mechanisms compress streams, fold completed sub-trajectories into abstractions, and use plans or structured world representations to guide future actions and revise strategies.Source: §4.3.2 · pp. 45–46

Motivation

• Agentic memory dynamically constructs and updates its memory store while performing query-conditioned retrieval, unlike statically encoded knowledge or fixed databases.This adaptive capability supports agent self-evolution and lifelong learning.Source: §5 · p. 46

Dynamic Memory Lifecycle

• The memory lifecycle extracts generalizable knowledge from reasoning traces and environmental feedback, fuses it with existing memories, and updates the memory base to adapt continuously and mitigate cognitive conflicts.The constructed memory base supports targeted retrieval from designated modules at precise moments to enhance reasoning.Source: §5 · p. 46
• The section decomposes dynamic memory management into three interacting processes—formation, evolution, and retrieval—that support adaptive, self-evolving agent behavior.Figure 8 provides a holistic illustration of these lifecycle processes.Source: §5 · p. 46

• The section presents Knowledge Distillation Structured Construction.Source: §1 · p. 47

3. Memory Retrieval

• Latent-representation parametric internalization is identified as a related memory mechanism.Source: §3 · p. 47
• Memory retrieval includes querying related memories.Source: §3 · p. 47
• Memory retrieval also considers existing similar memories.Source: §3 · p. 47

Memory Evolution

• Memory evolution integrates newly extracted memories with the existing repository through consolidation, updating, and forgetting operations.These operations keep the memory repository coherent and efficient as it changes over time.Source: §2 · p. 47

Operational Dynamics

• The memory lifecycle comprises formation, evolution, and retrieval, which respectively extract useful knowledge, refine the repository, and provide context-aware information for reasoning.Figure 8 presents these as the system’s three fundamental operational processes.Source: Figure 8 · p. 47, §2 · p. 47

Memory Evolution

• Memory evolution refines the memory base by consolidating correlated entries, resolving conflicts, and adaptively pruning information to preserve generalizability, coherence, and efficiency.Its purpose is to integrate newly formed memories with existing knowledge in changing environments.Source: §2 · p. 47

Operational Dynamics

• These processes form an interconnected cycle: reasoning outcomes and environmental feedback generate new insights for formation and refinements for memory evolution.The resulting system continuously learns from and responds to changing environments rather than operating as a static conditional generator.Source: §2 · p. 48

5.1 Memory Formation

• Memory formation encodes raw contexts into compact knowledge, reducing computational overhead, memory footprints, and degraded reasoning caused by lengthy, noisy, redundant inputs.Formation selectively extracts architectural memories to fulfill task-specific functions, and Table 7 compares methods by subtype, representation form, and mechanism.Source: §5.1 · p. 48
• The five formation strategies are complementary rather than mutually exclusive, allowing algorithms to combine operations and translate knowledge across representations.This flexibility connects formation choices to task requirements, memory functions, and architectural forms.Source: §5.1 · p. 48

5.1.1 Semantic Summarization

• Semantic summarization preserves global, high-level meaning while filtering redundancy, improving efficiency but sacrificing fine-grained factual precision.Incremental methods evolve summaries over time, whereas partitioned methods summarize semantic units separately; both face trade-offs involving drift, forgetting, discontinuity, or cross-partition dependencies.Source: §5.1.1 · p. 50

5.1.2 Knowledge Distillation

• Knowledge distillation extracts function-specific factual or experiential knowledge, preserving user and environmental facts or generalizing planning strategies from successful and failed trajectories.H2R addresses the lack of fine-grained insights by combining high-level planning reflections with step-wise execution insights.Source: §5.1.2 · p. 51

5.1.3 Structured Construction

• Structured construction organizes information into linked and layered topological representations, improving interpretability, retrieval efficiency, and multi-hop reasoning over complex dependencies.Entity-level and chunk-level paradigms support planar, dynamic, and hierarchical structures, but schema rigidity and extraction or maintenance costs remain limitations.Source: §5.1.3 · pp. 52–53

5.1.4 Latent Representation

• Latent representation encodes experiences directly as machine-native vectors, embeddings, or KV-caches, supporting dense semantic signals and unified multimodal alignment.Its principal limitation is opaqueness: latent memories are difficult for humans to debug, edit, or verify.Source: §5.1.4 · pp. 53–54

5.1.5 Parametric Internalization

• Parametric internalization consolidates external memories into model parameters, eliminating external retrieval overhead while supporting continual updates and intrinsic competence.The section distinguishes knowledge internalization from capability internalization; the former converts factual memories into parameters through model editing or lightweight adapters.Source: §5.1.5 · p. 54
• Knowledge internalization methods progressed from single-step edits in MEND to precise ROME, scalable MEMIT, and adapter-based CoLoR.Despite improved precision, scalability, and parameter efficiency, these methods remain vulnerable to off-target effects and catastrophic forgetting.Source: §5.1.5 · p. 54
• Parametric internalization embeds experiential knowledge—including solution schemas, strategic planning, and agentic skills—into the model’s parameter space.It learns from reasoning traces through supervised fine-tuning (Wei et al., 2022; Mukherjee et al., 2023) and preference optimization such as DPO and GRPO.Source: §5.1.5 · p. 54
• Memory Decoder integrates external RAG with parameterized training as a plug-and-play method that preserves the base model while eliminating external retrieval overhead.It combines external-RAG-like deployment with parameter-internalized inference speed.Source: §5.1.5 · p. 54
• Parameter-fused knowledge becomes effectively instinctive, enabling immediate responses without querying external memory.This represents memory’s ultimate consolidation through gradient-based fusion into model weights and mimics biological long-term potentiation.Source: §5.1.5 · pp. 54–55
• Parametric internalization faces catastrophic forgetting and high update costs, limiting its practicality for continual memory formation.Unlike external memory, it is difficult to modify or remove precisely without unintended side effects.Source: §5.1.5 · p. 55

5.2 Memory Evolution

• Memory Evolution combines consolidation, updating, and forgetting to maintain the memory repository’s generalization, accuracy, timeliness, and efficiency.Figure 9 presents these mechanisms around a shared Memory Database, including local, cluster-level, and global consolidation; conflict resolution and parameter updates; and pruning by time, access frequency, or informational value.Source: §5.2 · p. 55, Figure 9 · p. 56

5.2.1 Consolidation

• Consolidation transforms fragmented short-term traces into generalized long-term knowledge through local merging, cluster-level fusion, and global integration.It improves cross-task generalization and reasoning structure but can smooth away outliers and rare exceptions, reducing sensitivity to specific events.Source: §5.2.1 · pp. 56–57

5.2.2 Updating

• Updating revises external memories or model parameters when new information conflicts with existing knowledge, enabling continuous self-correction without full retraining.External updates evolved toward temporally aware, delayed-consistency, and learned policies, while model editing directly modifies parameter space; incorrect updates risk knowledge degradation.Source: §5.2.2 · p. 58

5.2.3 Forgetting

• Forgetting removes outdated, redundant, or low-value information through time-based, frequency-based, and importance-driven criteria to control overload and preserve relevance.Aggressive pruning can erase rare but essential knowledge, so systems may avoid deletion when storage cost is not critical; MemEvolve instead evolves experiential knowledge and the memory architecture itself.Source: §5.2.3 · pp. 58–59

5.3 Memory Retrieval

• The retrieval pipeline integrates timing and intent, query construction, retrieval strategies, and post-retrieval processing to support efficient, context-aware knowledge access.Figure 10 organizes existing methods across these four stages, which respectively govern initiation, query formulation, search, and output refinement.Source: §5.3 · p. 60, Figure 10 · p. 61

5.3.1 Retrieval Timing and Intent

• Autonomous timing and intent let agents decide when to retrieve and which memory store to query, but missed retrieval can cause silent failures and hallucinated outputs.MemGen uses latent triggers based on rollout states, while H-MEM performs coarse-to-fine routing across memory hierarchies.Source: §5.3.1 · pp. 60–61

5.3.2 Query Construction

• Query construction uses decomposition or rewriting to align user inputs with memory indices, and targeted or planned queries can improve retrieval precision and reasoning performance.Agent KB (Tang et al., 2025d) generates sub-queries from observed student failures, while Auto-RAG (Kim et al., 2024a) selects between decomposition and rewriting under identical retrieval conditions.Source: §5.3.2 · p. 62

5.3.3 Retrieval Strategies

• Retrieval strategies span lexical, semantic, graph, generative, and hybrid paradigms, trading interpretability, semantic generalization, structural reasoning, scalability, and comprehensive coverage.Semantic retrieval is common but can introduce drift and spurious recall; graph methods support multi-hop and constrained reasoning, whereas generative retrieval is less scalable as corpora evolve.Source: §5.3.3 · pp. 63–64

5.3.4 Post-Retrieval Processing

• Post-retrieval processing re-ranks and filters noisy results, then aggregates and compresses relevant fragments into concise contexts tailored to tasks, temporal constraints, or agent roles.Methods include reinforcement-learned signal weighting, timestamp filtering, LLM-based evaluation, abstract summarization, query-focused extraction, and role-specific personalization.Source: §5.3.4 · pp. 64–65

6.1.1 Benchmarks for Memory / Lifelong / Self-Evolving Agents

• Memory-oriented benchmarks test constructing, maintaining, and exploiting explicit memories across dialogues, sessions, narratives, and sometimes multimodal signals, with Table 8 summarizing their evaluation settings.Examples stress user modeling, preference tracking, and conversational consistency under controllable simulated conditions.Source: §6.1.1 · p. 65, Table 8 · p. 66
• Lifelong-learning benchmarks evaluate continual acquisition, consolidation, updating, forgetting, transfer, and test-time adaptation as information and task distributions evolve over long horizons.Self-evolving benchmarks additionally assess reflection, memory editing, tool-augmented storage, policy improvement, and strategy refinement across episodes.Source: §6.1.1 · pp. 65–67

6.1.2 Other Related Benchmarks

• Embodied, web, tool-use, multitask, planning, software, research, and multimodal benchmarks indirectly evaluate memory through long-horizon state tracking, context management, cross-step consistency, and reuse of prior information.These environments complement explicit memory tests by embedding memory demands in realistic sequential interactions and complex workflows.Source: §6.1.2 · p. 67

6.2 Open-Source Frameworks

• Open-source memory frameworks increasingly support factual stores, experiential traces, multimodal memory, and hierarchical, graph-based, profile-based, or modular organizations, while evaluation coverage remains uneven.They range from agent-centric systems to general retrieval and memory-as-a-service backends, with some separating short- and long-term stores and others leaving behavior and evaluation to applications.Source: §6.2 · pp. 68–69, Table 9 · p. 68

7.1.1 Look Back: From Memory Retrieval to Memory Generation

• Generative memory shifts beyond retrieval by synthesizing context-specific representations, with future systems adapting granularity, integrating heterogeneous signals, and learning memory generation through reinforcement learning.Existing approaches either retrieve before generating refined memories or generate directly from context, interaction history, or latent states.Source: §7.1.1 · p. 69, §7.1.2 · pp. 69–70

7.2.1 Look-Back: From Hand-crafted to Automatically Constructed Memory Systems.

• Manual memory strategies offer low-cost interpretability and control but generalize poorly, motivating agents that autonomously construct, evolve, retrieve, and self-organize memory through explicit tool calls and adaptive hierarchies.Tool-based memory operations can make behavior more coherent, transparent, and contextually grounded, while self-evolving structures reduce dependence on hand-designed rules.Source: §7.2.1 · p. 70, §7.2.2 · p. 70

7.3.2 Future Perspective

• Fully reinforcement-learning-driven memory is projected as the next major stage, enabling agents to invent memory organizations and control memory formation, evolution, and retrieval end to end.The proposed trajectory moves from heuristic pipelines, through partial RL involvement, toward model-native architectures and control policies learned jointly.Source: Figure 11 · p. 71, §7.3.2 · p. 72

7.4.2 Future Perspective

• Multimodal memory must unify heterogeneous perceptual, action, and environmental signals, but no current system provides truly omnimodal support; future systems require cross-modal abstraction, reasoning, alignment, and temporal coherence.Visual and video memory are more developed than audio and other modalities, while unified-model memory also supports consistency across long generation horizons.Source: §7.4.1 · pp. 72–73, §7.4.2 · p. 73

7.5.2 Future Perspective

• Shared multi-agent memory is evolving from isolated or centralized repositories toward agent-aware, learning-driven collective representations conditioned on roles, expertise, trust, and long-horizon team performance.Future shared memory must also abstract across heterogeneous signals while preserving temporal and semantic coherence.Source: §7.5.2 · p. 73, §7.5.1 · p. 73

7.6.2 Future Perspective

• World-model memory is transitioning from passive data caching to active state simulation through dual fast–slow systems and policies that curate, summarize, and discard information by task relevance.This architecture pairs efficient state-space processing for immediate interaction with deliberative models or explicit memory for planning and consistency.Source: §7.6.2 · p. 74

7.7.1 Look-Back: From Trustworthy RAG to Trustworthy Memory

• Trustworthy memory must address persistent, user-specific risks beyond RAG, including privacy leakage, unsafe retention, weak access control, unverifiable forgetting, unauditable updates, explainability gaps, and collective privacy.Long-lived memory amplifies these risks across individual, multi-agent, and federated deployments.Source: §7.7.1 · p. 75

7.7.2 Future Perspective

• The section ultimately identifies privacy preservation, explainability, and hallucination robustness as three interlinked pillars for trustworthy memory architectures and algorithms.These pillars elevate trust from an auxiliary concern to a first-class design principle for agent memory.Source: §7.7.2 · p. 75, §7.7.1 · p. 75
• Trustworthy memory requires permissioned retention, encrypted or on-device storage, explainable access paths, hallucination safeguards, and auditable OS-like abstractions jointly managed by agents and users.Proposed safeguards include differential privacy, memory redaction, adaptive forgetting, uncertainty-aware retrieval, abstention, fallback to model priors, and multi-agent cross-checking.Source: §7.7.2 · pp. 75–76

7.8.1 Look Back

• Current agent memory mirrors human memory’s structural categories but mainly retrieves immutable tokens, lacking human-like distortion, abstraction, and dynamic remodeling of history.Contemporary systems pair capacity-limited context windows with external vector databases and organize interaction logs, world knowledge, and code-based skills into episodic, semantic, and procedural analogues.Source: §7.8.1 · p. 76

7.8.2 Future Perspective

• Future agents should add sleep-like offline consolidation to reorganize memories, replay experiences, distill schemas, forget redundant noise, and optimize indices without real-time latency.This direction draws on Complementary Learning Systems theory and addresses the limits of exclusive online updating.Source: §7.8.2 · p. 76
• Generative memory could replace explicit text retrieval with on-demand latent reconstruction, compacting episodic streams into parametric intuition while resolving the stability-plasticity dilemma.This shift would move agents from archiving data toward internalizing experience through periodic consolidation cycles.Source: §7.8.2 · p. 76

• The survey identifies token-level, parametric, and latent memory as three principal realizations with distinct trade-offs in representation, adaptability, and policy integration.These forms reflect rapid advances in contemporary agent memory research.Source: §8 · p. 76
• It replaces the coarse long-term versus short-term distinction with factual, experiential, and working memory, corresponding to knowledge retention, capability accumulation, and task-level reasoning.Source: §8 · pp. 76–77
• Memory is presented as an essential substrate for temporal coherence, continual adaptation, and long-horizon competence rather than merely an auxiliary storage mechanism.Source: §8 · p. 77
• Reinforcement learning integration, multimodal and multi-agent settings, and generative memory point toward fully learnable, adaptive, and self-organizing memory systems.These developments could enable sustained interaction, self-improvement, and principled reasoning over time.Source: §8 · p. 77
• The survey positions memory design as a central and open problem in developing robust, general, and enduring artificial intelligence.It is intended as a foundation and reference for future researchers and practitioners.Source: §8 · p. 77