PaperlayerView exact revision on arXiv

Source-linked AI summary

Rethinking Memory in LLM based Agents: Representations, Operations, and Emerging Topics

Yiming Du, Wenyu Huang, Danna Zheng, Zhaowei Wang, Sebastien Montella, Mirella Lapata, Kam-Fai Wong, Jeff Z. Pan

arXiv:2505.00675v3cs.CL

TL;DR

Existing surveys do not provide a unified framework for the operations governing LLM-agent memory. This survey classifies memory types and operations, identifies research topics, and highlights gaps between retrieval and generation.

  • Problem

    Existing surveys lack a unified framework for organizing and evolving the memory operations of LLM-based agents.

  • Method

    The survey classifies agent memory into parametric and contextual types and defines six core memory operations.

  • Results

    Retrieval performance can exceed 90 Recall@5 while generation F1 lags by over 30 points, showing that retrievability does not guarantee effective generation.

  • Takeaways & Limitations

    The taxonomy highlights temporal information as important for long-horizon memory, with 60% of surveyed models encoding it.

  • Takeaways & Limitations

    Multi-source memory can lack consistency checks or source attribution, leading to hallucinations and factual drift.

Abstract

from arXiv · show

Memory is fundamental to large language model (LLM)-based agents, but existing surveys emphasize application-level use (e.g., personalized dialogue), while overlooking the atomic operations governing memory dynamics. This work categorizes memory into parametric (implicit in model weights) and contextual (explicit external data, structured/unstructured) forms, and defines six core operations: Consolidation, Updating, Indexing, Forgetting, Retrieval, and Condensation. Mapping these dimensions reveals four key research topics: long-term, long-context, parametric modification, and multi-source memory. The taxonomy provides a structured view of memory-related research, benchmarks, and tools, clarifying functional interactions in LLM-based agents and guiding future advancements. The datasets, papers, and tools are publicly available at https://github.com/Elvin-Yiming-Du/Survey_Memory_in_AI.

1 Introduction

This survey addresses the lack of a unified operational framework for LLM-agent memory by organizing memory representations, types, and six core operations. It further identifies four research frontiers and provides systematic evidence, practical resources, and implementation guidance for memory-enabled agents.

  • 1 Introduction: The survey distinguishes parametric memory, encoded implicitly in model parameters, from contextual memory, which stores explicit structured or unstructured external information.It also considers long-term memories such as multi-turn dialogue and short-term contexts such as KV-cache and current dialogue history.
  • 1 Introduction: It formalizes six memory operations across encoding, evolving, and adapting: consolidation, indexing, updating, forgetting, retrieval, and condensation.These operations integrate and organize knowledge, modify or remove memory, access relevant information, and reduce size while preserving key content.
  • 1 Introduction: The framework organizes four research frontiers: long-term memory, long-context memory, parametric memory modification, and multi-source memory.These span temporal persistence and personalization, efficient extended-context handling, internal knowledge editing and unlearning, and integration across textual or multimodal sources [102, 281].
  • 1.1 Research Methodology: The methodology analyzes 37 foundational or representative seed papers, uses expert annotation and iterative discussion to define a taxonomy, and collects over 30,000 papers from 2022–2025 venues.Papers scoring ≥8 out of 10 according to taxonomy-aligned task definitions were retained.
  • 1.1 Research Methodology: R^2 = 0.97: the log-log regression used for the Relative Citation Index achieved the best fit and supported time-normalized impact comparisons across papers and years.The resulting semantic-relevance and RCI pipeline provides a reproducible basis for analyzing research progress and topic-specific impact across the four areas.
  • 1.2 Contribution and Structure: The survey contributes a conceptual foundation for memory lifecycles and a practical reference covering tools, products, benchmarks, deployment scenarios, emerging trends, and open challenges.Its structure proceeds from representations and operations, through high-impact topics and applications, to human–agent comparisons and future directions.

2 Memory Foundations

This section establishes foundations for understanding agent memory through its representations, timescales, functional types, and core operations. It examines how memory is stored, updated, forgotten, retrieved, and condensed across parametric and contextual mechanisms.

  • 2 Memory Foundations: Memory foundations integrate representation, timescale, functional type, and operations to explain what memory is, how long it persists, its role, and how it evolves.Together, these dimensions connect memory structure, function, and dynamics in human cognition and artificial intelligence.
  • 2.1 Memory Representation: Parametric memory stores knowledge implicitly in model weights for fast, context-free factual retrieval, whereas contextual memory stores explicit external information in unstructured or structured forms.Parametric memory is persistent but difficult to update selectively and lacks transparency; contextual memory can include text, images, audio, videos, embeddings, knowledge graphs, tables, and ontologies.
  • 2.2 Memory Timescale: Memory timescale distinguishes long-term persistence across extended interactions from short-term storage for immediate reasoning and decision-making.Long-term memory supports continuity, personalization, and knowledge grounding, while short-term memory typically uses the KV cache or current context window and can be consolidated into long-term storage.
  • 2.2 Memory Timescale: Short-term memory supports long-context reliability by mitigating hallucinations, the “lost in the middle” problem, multi-turn error accumulation, and unreliable multi-turn tool use.These functions are especially important in ultra-long contexts and multi-turn interactions.
  • 2.3 Memory Functional Type: Functional memory comprises episodic, semantic, procedural, and working memory, covering experiences, facts, learned skills, and active information manipulation.Episodic memory complements parametric knowledge with evolving experiences; semantic memory is relatively stable but can blur with episodic knowledge in LLMs; procedural memory supports automated task behavior; and working memory assembles inputs for reasoning, planning, and goal-directed action.
  • 2.4 Memory Operations: Dynamic memory requires operations that govern information’s lifecycle and use during interaction, grouped into memory encoding, evolving, and adapting.The supplied operations discussion introduces these categories and motivates operations beyond static storage.
  • 2.4.1 Memory Encoding.: Memory encoding transforms information into storable, retrievable representations through consolidation and indexing.Consolidation converts short-term experiences into persistent forms such as parameters, graphs, or knowledge bases, while indexing creates access codes and temporal or relational paths for scalable retrieval.
  • 2.4.2 Memory Evolving.: Memory evolving is introduced as dynamic change over time through complementary updating and forgetting processes.This part identifies updating and forgetting as the two processes of memory evolution, without detailing their mechanisms further.
  • 2.4.2 Memory Evolving.: Memory evolving updates existing representations with new knowledge and reorganizes contextual content through summarization, pruning, or refinement [13].Parametric updating typically uses locate-and-edit mechanisms, whereas contextual updating replaces or reorganizes outdated content.
  • 2.4.2 Memory Evolving.: Forgetting selectively suppresses outdated, irrelevant, or harmful content through parametric unlearning or contextual deletion and semantic filtering [119] [157] [296].These operations help maintain memory efficiency, but attackers can poison memory, leaving corrupted fragments that persist undetected and trigger malicious actions.
  • 2.4.3 Memory Adapting.: The approach encompasses retrieval and compression; retrieval identifies and accesses relevant information from memory in response to inputs, supporting downstream tasks, while threshold 𝜏 marks relevant fragments from multiple sources or modalities.
  • 2.4.3 Memory Adapting.: Condensation enables efficient context usage under limited context windows by retaining salient information and discarding redundancies at compression ratio α, through pre-input or post-retrieval compression.Pre-input compression scores, filters, or summarizes full-context inputs to fit context constraints, while post-retrieval compression reduces retrieved content through contextual or parametric compression.

3 From Operations to Key Research Topics

This section maps core memory operations to four research topics—long-term memory, long-context memory, parametric-memory modification, and multi-source memory—covering their methods, evaluation, and management challenges. It highlights issues in retrieval, compression, updating, forgetting, personalization, integration, and long-horizon maintenance, including the need for dynamic, operation-aware benchmarks.

  • 3 From Operations to Key Research Topics: The framework analyzes how real-world systems manage memory operations across four key research topics, using the time-adjusted Relative Citation Index to identify influential work.Figure 2 maps core memory operations to the four topics, while RCI surfaces emerging trends and enduring contributions.
  • 3.1.1 Memory Encoding.: Long-term memory encodes raw experiences through consolidation and indexing, preserving contextual, episodic, and semantic information while integrating relational, metadata, and temporal retrieval signals.Consolidation spans dialogue, reasoning, parametric, event-level, and knowledge-level representations, whereas indexing uses graph-based, signal-enhanced, and timeline-based paradigms.
  • 3.1.2 Memory Evolving.: Memory evolving uses updating and forgetting to incorporate new knowledge, correct or remove outdated content, and maintain accuracy, efficiency, relevance, safety, and privacy.Active forgetting targets sensitive, harmful, or private consolidated information, but must avoid collateral damage to valid knowledge; updates must preserve prior knowledge and factual or stylistic consistency.
  • 3.1.3 Memory Adapting.: Memory adapting retrieves relevant entries and condenses them into compact contexts for grounded generation, progressing from static integration toward agentic, self-improving context restructuring.Retrieval includes query-, memory-, and event-centered paradigms, while ACE [361] enables autonomous prioritization and restructuring; noise and earlier inaccuracies remain threats to reliability.
  • 3.1.4 Personalization.: Personalization faces data sparsity, privacy, changing preferences, and a personalization-generalization trade-off, using model-level adaptation or external memory augmentation.Model-level methods encode user preferences through fine-tuning, lightweight updates, latent persona representations, prefix encoders, or modular adapters.
  • 3.1.4 Personalization.: Personalization retrieves user-specific information from external repositories using structured, unstructured, or hybrid memories, but often treats long-term memory as a passive buffer.These systems scale across profiles, knowledge graphs, dialogue histories, and persistent cross-session stores, while leaving proactive planning and decision-making underused.
  • 3.1.5 Discussion.: Current benchmarks emphasize retrieval and post-retrieval generation while overlooking consolidation, updating, forgetting, selective retention, temporal continuity, and user adaptation.The section calls for benchmarks spanning parametric, contextual unstructured, and structured memory with dynamic evaluation of reliability, temporal reasoning, and multi-session consistency.
  • 3.1.5 Discussion.: Recall@5 exceeds 90 on 2Wiki and MemoryBank, yet generation F1 lags by over 30 points, showing that retrievability does not ensure effective generation.Compact memories support generation better than verbose entries, while longer temporal distances, retrieval noise, and multilingual differences further impair memory-grounded generation.
  • 3.2 Long-context Memory: Long-context memory research addresses extensive multi-source external memory through memory compression and retrieval for efficient processing.Compression methods operate on context tokens or KV caches, spanning soft and hard prompt compression, eviction, summarization, hybrid approaches, and reinforcement-learning-based active management.
  • 3.2.1 Memory Compression.: KV-cache eviction reduces quadratic memory growth in extremely long contexts by removing unnecessary key-value pairs through fixed-pattern or adaptive selection.Static methods use predefined patterns, whereas dynamic methods select cache entries according to the query or model behavior during inference.
  • 3.2.1 Memory Compression.: Memory compression reduces long-context processing load, while retrieval methods identify key information and selectively load essential KV caches for faster inference.Compression methods include merging or low-rank representations, quantization, and software–hardware co-design; retrieval uses graph, token, fragment, training-based, vector-memory, or query-aware KV-cache selection.
  • 3.2.3 Discussion.: Compression methods balance memory cost against performance degradation: eviction achieves higher compression but greater information loss, while storage optimization preserves more performance at a memory cost.The survey identifies this balance as a central challenge for both parametric KV-cache and contextual-memory compression.
  • 3.2.3 Discussion.: Long-context models can exhibit “lost in the middle” failures in key-value retrieval, limitations in multimodal contexts, failures when reasoning from contextual memory, and degraded generation when irrelevant retrieved information misleads them.
  • 3.2.3 Discussion.: KV-cache storage optimization generally offers the best effectiveness–efficiency trade-off, whereas eviction is more flexible but less effective and contextual compression performs comparatively poorly.Figure 6 identifies storage optimization with x markers, eviction with ∇ markers, and contextual compression with Δ markers; LLMLingua2 exemplifies the latter comparison.
  • 3.3 Parametric Memory Modification: Parametric memory modification comprises editing, unlearning, and continual learning to adapt encoded knowledge, remove targeted information, and mitigate catastrophic forgetting.The section reviews these categories systematically as approaches to dynamically adapting stored model knowledge.
  • 3.3.1 Editing.: Parametric editing updates specific knowledge without full retraining through locating-then-editing, meta-learning, prompt-based steering, or additional parametric modules.Locating-then-editing methods identify where facts are stored before modifying them, while other methods predict, prompt, or augment behavior without directly changing original weights [52] [107] [203] [205].
  • 3.3.2 Unlearning.: Parametric unlearning selectively removes targeted memories while preserving unrelated knowledge through added modules, prompt manipulation, targeted deactivation, or objective-based training.These methods seek a balance between forgetting efficiency and precision [118, 30, 173, 231, 119, 277, 317, 120, 185, 304, 343].
  • 3.3.3 Continual Learning.: Continual learning supports long-term parametric-memory persistence by mitigating catastrophic forgetting through regularization-based and replay-based methods.Regularization constrains updates to important weights, preserving vital parametric memory; examples include TaSL [73] and SELF-PARAM.
  • 3.3.3 Continual Learning.: Continual-learning methods use constraints or replay to incorporate knowledge without full retraining, while LSCS incrementally acquires and consolidates memory through real-time interaction.DSI++ uses generative memory and pseudo queries to maintain retrieval performance without full retraining.
  • 3.3.4 Discussion.: Most editing methods support only 1,000–5,000 sequential edits, while MemoryLLM supports up to 650k; non-prompt methods typically target models ≤20B and face higher computational costs.These scaling patterns motivate balancing model size with editing or unlearning effectiveness and efficiency.
  • 3.3.4 Discussion.: Prompt-based methods perform strongly across CounterFact, ZsRE, and ToFU, whereas meta-learning methods generally underperform; ZsRE suffers from lower specificity, and ToFU may be insufficiently challenging.Figure 10 compares SOTA methods across the three benchmarks; the discussion calls for improved specificity and new unlearning benchmarks.
  • 3.4.1 Cross-textual Integration.: Multi-source reasoning combines parametric memories with structured and unstructured external knowledge, but unified reasoning remains difficult because systems must resolve conflicts and calibrate trust.Models may favor internal knowledge over retrieved content, motivating source attribution and trust calibration; existing methods often remain static or single-source.
  • 3.4.2 Multi-Modal Coordination.: Multimodal memory systems align and retrieve heterogeneous inputs, yet struggle with long-term updates, cross-modal retrieval, and reasoning across audio and sensorimotor signals.M3 unifies storage and reasoning across diverse data types, enabling dynamic updates and reasoning-driven retrieval for long-term memory management.
  • 3.4.3 Discussion.: Multi-source systems are shifting from static retrieval toward dynamic, context-sensitive memory supporting temporally grounded, cross-source reasoning.This transition is reflected in retrieval-generation loops, collaborative agents, and active memory maintenance.
  • 3.4.3 Discussion.: Cross-textual integration spans symbolic versus unstructured sources and explicit queries versus neural reasoning, but most systems remain disconnected from real-time inference.Symbolic approaches offer transparency but limited open-domain scalability, whereas newer methods combine neural retrieval, fusion, and chain-of-thought reasoning.
  • 3.4.3 Discussion.: Conflict resolution across heterogeneous sources remains a major challenge because retrieved and parametric content can lack consistency checks or source attribution, causing hallucinations and factual drift.Multi-step conflict resolution [299] and epistemic calibration [328] are promising but currently lack scalability.
  • 3.4.3 Discussion.: 60% of surveyed models encode temporal information, while multimodal memory advances through fusion, retrieval, and temporal modeling for extended-interaction reasoning.Strategies include joint or prompt-level fusion, identifier-based and graph-based integration, and temporal graphs or time-aware attention.
  • 3.4.3 Discussion.: Operational control is becoming essential as newer agents add updating, indexing, compression, and self-maintenance, shifting memory from passive querying to active refinement.WorldMem compresses multimodal logs, while E-Agent dynamically updates internal memory for long-horizon planning.
  • 3.4.3 Discussion.: Cross-textual reasoning dominates publication volume, fusion research led by CLIP [237] has the highest citation impact, and dynamic retrieval and conflict resolution remain underexplored.The field is consequently moving toward deeper, operation-aware, temporally structured architectures with explicit conflict detection, attribution, and resolution.

4 Memory In Practice

Memory in practice spans products, development tools, and infrastructure, whose bidirectional interaction supports increasingly capable agent memory. Products remain early in integration, motivating scalable and adaptive memory mechanisms.

  • 4 Memory In Practice: Memory augmentation forms an interdependent hierarchy in which product requirements drive tool design, tools constrain infrastructure, and infrastructure enables richer product capabilities.Bridging these layers clarifies the technical and conceptual frontiers of agent memory mechanisms.
  • 4.1 Products: Agent products are categorized by dominant memory types and application focus: personal agents primarily use contextual memory, task-oriented agents rely on it, and multimodal agents unify both.
  • MemOS [168]: Products remain at an early stage of memory integration because memory operations are largely prompt-based rather than dynamically managed.Limitations include hallucination, privacy concerns, memory drift, restricted multimodal reasoning, and insufficient persistent task- or project-level coordination.
  • 4.2 Development Tools: Development tools provide configurable pipelines and service layers that abstract memory operations, enabling multimodal, persistent, updatable, and temporally organized memory modules.Frameworks include Graphiti [92], LlamaIndex, LangChain, LangGraph, EasyEdit, CrewAI [66], MemU, and Letta; memory-layer systems include Mem0, Zep [239], Memary, MemOS [168], and Memobase.
  • 4.3 Infrastructure: Infrastructure combines persistent graph and vector storage, sparse and dense retrieval, LLM-guided lifecycle operations, and acceleration for high-throughput memory processing.These components support dynamic updating, targeted forgetting, ultra-long contexts, and massive interaction histories.

5 The Cognitive Gap between Biological and Agent Memory

The section argues that agent memory remains fundamentally less dynamic than biological memory, which hierarchically encodes, evolves, and reconstructs experience. It identifies gaps in constructive encoding, long-term internalization, adaptive retrieval, semantic coherence, ownership, and efficiency.

  • 5 The Cognitive Gap between Biological and Agent Memory: Agent memory remains focused on static persistence, lacking the dynamic sophistication of biological memory in encoding and evolving experience.Table 3 summarizes the key differences between human and agent memory.
  • 5 The Cognitive Gap between Biological and Agent Memory: Agents typically record verbatim data or statically parameterize weights, accumulating fragmented traces instead of constructing a synthesized self-model.This raw-data dependence prevents pruning noise at entry, while pre/post-training [171] remains a discrete process that does not produce adaptive cognitive structures.
  • Memory Evolving: Long-context scaling and retrieval bypass semantic internalization, reducing signal-to-noise ratio and increasing computational cost while retrieval without reconsolidation struggles to form coherent causal representations.The ’lost-in-the-middle’ phenomenon [179] illustrates this limitation; active rewriting methods such as AgentFold [344] and FoldGRPO [264] motivate real-time reconstruction of internal state.
  • Memory Evolving: The proposed direction is to move beyond passive retrieval toward active thinkers that reconstruct internal states in real time and prioritize semantic density for bio-inspired efficiency.This shift is presented as necessary for adapting to environmental dynamics while addressing the scalability costs of silicon-based memory.
  • Memory Evolving: Unlike holistic, associative, and metabolically frugal human memory, agent memory uses disconnected representations, is replicable and broadcastable, and incurs computational and environmental costs.Agent formats such as documents, graphs, and vector embeddings prioritize local pattern matching over global semantic coherence; human memory operates at approximately ∼20W.

6 Open Challenges and Future Directions

This section identifies open challenges in core memory topics and proposes future directions, including broader perspectives that extend memory systems’ capabilities and theoretical grounding. Together, these discussions provide a roadmap toward reliable, interpretable, and adaptive memory in AI.

  • 6 Open Challenges and Future Directions: The section surveys open challenges in core memory topics and broader directions including biologically inspired models, lifelong learning, multi-agent memory, and unified memory representation.These perspectives are presented as extending memory systems’ capabilities and theoretical grounding.
  • 6.1 Topic-Specific Directions: Guided by RCI analysis and observed trends, the section outlines key challenges and emerging demands shaping future memory research.The focus is on designing memory-centric AI that addresses core limitations.

Unified evaluation is needed to address consistency, personalization, and temporal reasoning in long-term

Existing benchmarks inadequately evaluate memory operations in dynamic, multi-session settings, contributing to retrieval–generation mismatch and motivating unified evaluation for temporal reasoning, personalization, and adaptive memory management. Long-context systems must balance computational efficiency, modeling fidelity, selective integration, and stability, increasingly through active context-management mechanisms.

  • Benchmarks rarely assess consolidation, updating, retrieval, and forgetting dynamically across sessions, leaving retrieved content outdated, irrelevant, or misaligned with generation.This gap contributes to the retrieval–generation mismatch caused by poor memory maintenance.
  • Addressing long-term memory requires temporal reasoning, structure-aware generation, retrieval robustness, personalized reuse, and adaptive management across sessions.
  • Long-context Processing: Efficiency vs. Expressivity: Long-context processing intensifies the trade-off between computational cost and modeling fidelity; KV cache compression and recurrent memory reuse improve efficiency but risk information loss or instability.
  • Long-context Processing: Efficiency vs. Expressivity: Multi-source and multi-modal reasoning requires selective context integration, source differentiation, and attention modulation, motivating RL-based frameworks for active context management.These mechanisms aim to balance contextual bandwidth with task relevance and stability.

While promising, parametric memory modification requires further research to improve control, erasure,

Parametric memory modification remains limited by insufficient specificity, scalability, and realistic evaluation, while broader memory research spans procedural skills, dynamic sharing, lifelong learning, multi-agent and multimodal coordination, biological inspiration, retrieval, temporal structure, unified representation, and safety.

  • While promising, parametric memory modification requires further research to improve control, erasure,: Parametric editing often lacks specificity, scales poorly beyond thousands of edits or models over 20B parameters, and relies on benchmarks such as TOFU that may be too simple.Future work should improve realistic benchmarks, efficiency, and the integration of editing, unlearning, and continual learning.
  • While promising, parametric memory modification requires further research to improve control, erasure,: Multi-source memory must resolve redundancy, inconsistency, ambiguity, temporal misalignment, conflicting semantics, and missing cross-modal attribution through conflict resolution, temporal grounding, provenance tracking, indexing, and compression.These capabilities support scalable and interpretable memory across structured knowledge, unstructured histories, and multimodal signals.
  • 6.2 Broader Perspectives: Parametric Sharing exchanges memory through model-native representations [21], such as dynamic adapters or specialized memory layers [383], preserving semantic nuances and enabling collaborative parametric intelligence.Standardized neural memory protocols and cross-model weight alignment remain future requirements for heterogeneous agents.
  • 6.2 Broader Perspectives: Lifelong learning should manage real-time environment streams [361] while mitigating catastrophic forgetting and enabling rapid adaptation [73], using meta-learned memory value and selective internalization [277].Personalized parameter spaces, including evolving LoRA or embeddings, can separate personal traces from general knowledge and support longitudinal, habit-aware companions.
  • 6.2 Broader Perspectives: Multi-agent and multimodal memory require sharing, alignment, conflict resolution, consistency, unified semantic spaces, and effective retrieval to support coordination, reasoning, and long-term collaboration.Current multimodal systems remain preliminary, with weak multi-turn reasoning and data misalignment.
  • 6.2 Broader Perspectives: Biological memory inspires dual-memory architectures, synaptic consolidation, experience replay, reconsolidation, bounded capacity, and hierarchical organization for update-aware recall, efficient storage, and context-sensitive generalization.These approaches address the stability–plasticity dilemma by separating fast-changing episodic experiences from slowly integrated long-term memory.
  • 6.2 Broader Perspectives: Open directions include selective retrieval from model parameters, efficient integration of latent memory, long-term spatio-temporal reasoning, and joint indexing mechanisms for memory consolidation and retrieval.

7 Conclusions · Appendix · A GPT-based Pipeline Selection

The survey organizes agent memory by type, operation, and functional perspective, then identifies four research topics and provides benchmarks and tool guidance. Its appendix describes a GPT-based pipeline for filtering over 30,000 papers by abstract–task alignment.

  • 7 Conclusions: The survey classifies agent memory into parametric and contextual types and maps its operations to encoding, evolving, and adapting.
  • 7 Conclusions: Its functional perspectives—episodic, semantic, procedural, and working memory—clarify how memory supports reasoning, personalization, and collaboration.
  • 7 Conclusions: The survey analyzes long-term memory, long context memory, parametric modification, and multi-source memory, highlighting progress, challenges, and future pathways.
  • 7 Conclusions: It also offers practical benchmarks and tool guidance for industry.
  • A GPT-based Pipeline Selection: The GPT-based pipeline pairs each paper abstract with a predefined task definition and scores alignment from 1–10, retaining papers scoring ≥8, as summarized in Table 4.
  • A GPT-based Pipeline Selection: GPT-4o-mini serves as the scoring backbone for cost-effective, sufficiently accurate zero-shot relevance estimation across a corpus exceeding 30,000 papers.The exact evaluation prompt is illustrated in Figure 18.

B Relative Citation Index

This section defines a Relative Citation Index (RCI) to identify impactful works by comparing observed citations with age-adjusted expectations. Using 3,932 relevant papers, it selects a log-log age-citation model because it best fits median citation patterns and behaves intuitively for newly released papers.

  • Data collection: The analysis collects papers from 2022–2025 at ACL, NAACL, EMNLP, NeurIPS, ICML, and ICLR, then uses GPT relevance scores to reduce research-area bias.Papers scoring at least 8 on relevance to the paper’s four topics are retained.
  • Data collection: After processing, 3,932 valid papers provide the data for estimating the age-citation model parameters.Papers lacking publication dates use their first conference day as the publication date, while citations are collected through the Semantic Scholar API.
  • Model selection: The log-log regression model best fits citation distributions across publication age, nearly matching median citations in the 3,932-paper dataset.Figure 15 shows citation distributions by age and the expected citation curve.
  • RCI computation: The selected model is used to estimate each paper’s expected citations according to publication age before computing its RCI.The model’s reported estimates are β̂ = 1.878 and α̂ = 1.297.
  • RCI interpretation: An RCI of at least 1 indicates citations above age-adjusted expectations, whereas lower RCI indicates the paper is under-cited relative to expectations.Higher RCI values indicate higher research impact.

C RCI-Driven Analysis of Topic Impact

RCI and publication-volume trends show that LLM advances substantially reshaped memory research, especially long-context and parametric memory. Long-term memory’s influence increased steadily, while other topics’ median RCI values declined after 2023 but remained comparable to pre-2022 levels.

  • C RCI-Driven Analysis of Topic Impact: In 2023, publication quantity and quality surged for long-context and parametric memory, suggesting these areas were directly shaped by LLM advances.Figure 16 summarizes median RCI distributions across topics by year.
  • C RCI-Driven Analysis of Topic Impact: Long-term and multi-source memory maintained relatively stable average impact levels, indicating continued activity without disruptive or field-defining work during that period.The comparison concerns the period surrounding the 2023 surge in LLM-related research.
  • C RCI-Driven Analysis of Topic Impact: Long-context grew from one of the least represented topics before 2022 to the most prominent by 2024, while long-term memory’s RCI increased steadily.These trends coincided with notable publication growth across all topics.
  • C RCI-Driven Analysis of Topic Impact: Other topics’ median RCI values declined after 2023, although their influence remained comparable to levels seen before 2022.Collectively, the trends underscore large models’ role in catalyzing progress across memory-related research, especially long-context and parametric memory.

D Chord Analysis of Interactions Among Memory Types, Operations, Topics, and Venues

The section uses chord-based analysis to examine memory research from two perspectives: interactions among memory types, operations, and topics, and their distribution across major ML and NLP conference venues.

  • D Chord Analysis of Interactions Among Memory Types, Operations, Topics, and Venues: Chord-based analysis is used to study memory research.The analysis covers both conceptual interactions and venue distribution.
  • D Chord Analysis of Interactions Among Memory Types, Operations, Topics, and Venues: The first perspective examines interactions among memory types, operations, and topics.
  • D Chord Analysis of Interactions Among Memory Types, Operations, Topics, and Venues: The second perspective examines how memory research is distributed across major ML and NLP conference venues.

D.1 Memory Interactions Across Types, Operations, and Topics

Across 132 method-focused papers, research concentrates on parametric and unstructured memory, especially compression, retrieval, forgetting, and updating, while structured and multi-source memory and indexing remain comparatively underexplored. The topic mapping links long-context and long-term work mainly to unstructured-memory operations, whereas parametric modification centers on parameter adaptation.

  • Research predominantly targets parametric and contextual unstructured memory, while contextual structured memory remains relatively underexplored because LLMs perform less effectively on structured inputs.
  • Compression and retrieval are the most studied operations, whereas indexing receives limited attention; consolidation commonly stores knowledge in model parameters or fixed external formats.
  • Updating and forgetting mainly support parametric knowledge editing and unlearning, but opaque model internals leave these operations at an early exploration stage.
  • Parametric modification centers on parametric memory, while long-context research emphasizes compression and retrieval, long-term research emphasizes consolidation, compression, and retrieval, and multi-source work remains limited.
  • Future work should integrate structured with unstructured memory, resolve conflicts across heterogeneous sources, and develop indexing methods tailored to LLM reasoning and interaction.

D.2 Memory Interactions Across Conference Venues · INSPIRED · Memory

The survey finds distinct operation and topic emphases across ML and NLP conference venues, while organizing evaluation datasets, methods, and tools across memory topics and abstraction levels.

  • D.2 Memory Interactions Across Conference Venues: Compression, forgetting, and updating appear more often in ML conferences, whereas retrieval and consolidation are more common in NLP conferences, as shown in Figure 20.The authors interpret the former operations as more theoretically exploratory and the latter as more practically grounded, leaving compression, forgetting, and updating with substantial real-world translation potential.
  • D.2 Memory Interactions Across Conference Venues: Indexing remains underrepresented in both ML and NLP venues, partly because it frequently co-occurs with retrieval and current vector-based approaches offer few novel variations.The survey notes that existing vector-based indexing approaches are relatively uniform.
  • D.2 Memory Interactions Across Conference Venues: Long-term memory and multi-source memory are more prevalent in NLP venues, while long-context memory and parametric modification appear more often in ML venues.The distribution is attributed to differing application- and theory-oriented emphases, with multi-source challenges arising during real-world applications and system integration.
  • INSPIRED: The survey catalogs datasets for long-term, long-context, parametric, and multi-source memory evaluation, including modalities, operations, task types, sources, and reasoning attributes.Tables 5–8 cover long-term, long-context, parametric, and multi-source evaluation datasets, respectively.
  • INSPIRED: The survey organizes methods for long-term memory, personalization, long-context compression and retrieval, parametric editing, unlearning, continual learning, and cross-textual integration.Tables 9–17 summarize method properties such as training-free operation, retrieval, full-context access, batch or sequential editing, task boundaries, and source types.
  • INSPIRED: The method overview also distinguishes continual-learning settings and cross-textual integration through task-boundary, task-setting, source-type, and source-dataset attributes.Tables 15–17 cover parametric modification in continual learning and multi-source memory in cross-textual integration.
  • Memory: Memory tools are classified at component, framework, application-layer, and product levels, with taxonomies covering operations, functions, inputs and outputs, example uses, and source types.Tables 18–21 provide the corresponding tool overviews, including product-level tools for memory utilization.
Loading 2505.00675v3…