Memory in Large Language Models: Mechanisms, Evaluation and Evolution

2025/09/23 by Zhang, Dianxing, Li, Wendong, Song, Kani +4 · 2 citations
#Artificial Intelligence (cs.AI) #FOS: Computer and information sciences

paper · doi:10.48550/arxiv.2509.18868

Abstract

Under a unified operational definition, we define LLM memory as a persistent state written during pretraining, finetuning, or inference that can later be addressed and that stably influences outputs. We propose a four-part taxonomy (parametric, contextual, external, procedural/episodic) and a memory quadruple (location, persistence, write/access path, controllability). We link mechanism, evaluation, and governance via the chain write -> read -> inhibit/update. To avoid distorted comparisons across heterogeneous setups, we adopt a three-setting protocol (parametric only, offline retrieval, online retrieval) that decouples capability from information availability on the same data and timeline. On this basis we build a layered evaluation: parametric (closed-book recall, edit differential, memorization/privacy), contextual (position curves and the mid-sequence drop), external (answer correctness vs snippet attribution/faithfulness), and procedural/episodic (cross-session consistency and timeline replay, E MARS+). The framework integrates temporal governance and leakage auditing (freshness hits, outdated answers, refusal slices) and uncertainty reporting via inter-rater agreement plus paired tests with multiple-comparison correction. For updating and forgetting, we present DMM Gov: coordinating DAPT/TAPT, PEFT, model editing (ROME, MEND, MEMIT, SERAC), and RAG to form an auditable loop covering admission thresholds, rollout, monitoring, rollback, and change audits, with specs for timeliness, conflict handling, and long-horizon consistency. Finally, we give four testable propositions: minimum identifiability; a minimal evaluation card; causally constrained editing with verifiable forgetting; and when retrieval with small-window replay outperforms ultra-long-context reading. This yields a reproducible, comparable, and governable coordinate system for research and deployment.

Citations

• Diverse And Private Synthetic Datasets Generation for RAG evaluation: A multi-agent framework
• Memento: Fine-tuning LLM Agents without Fine-tuning LLMs
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions
• BYOKG-RAG: Multi-Strategy Graph Retrieval for Knowledge Graph Question Answering
• Controlled Retrieval-augmented Context Evaluation for Long-form RAG
• MemBench: Towards More Comprehensive Evaluation on the Memory of LLM-based Agents
• CentroidKV: Efficient Long-Context LLM Inference via KV Cache Clustering
• GaRAGe: A Benchmark with Grounding Annotations for RAG Evaluation
• RARE: Retrieval-Aware Robustness Evaluation for Retrieval-Augmented Generation Systems
• How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior
• EnronQA: Towards Personalized RAG over Private Documents
• RAG Without the Lag: Interactive Debugging for Retrieval-Augmented Generation Pipelines
• Retrieval-Augmented Generation with Conflicting Evidence
• QE-RAG: A Robust Retrieval-Augmented Generation Benchmark for Query Entry Errors
• HoH: A Dynamic Benchmark for Evaluating the Impact of Outdated Information on Retrieval-Augmented Generation
• Quantifying Memorization and Parametric Response Rates in Retrieval-Augmented Vision-Language Models
• InfiniteHiP: Extending Language Model Context Up to 3 Million Tokens on a Single GPU
• MedRAG: Enhancing Retrieval-augmented Generation with Knowledge Graph-Elicited Reasoning for Healthcare Copilot
• TReMu: Towards Neuro-Symbolic Temporal Reasoning for LLM-Agents with Memory in Multi-Session Dialogues
• Episodic Memories Generation and Evaluation Benchmark for Large Language Models
• Visual RAG: Expanding MLLM visual knowledge without fine-tuning
• ASTRID -- An Automated and Scalable TRIaD for the Evaluation of RAG-based Clinical Question Answering Systems
• RAG-Check: Evaluating Multimodal Retrieval Augmented Generation Performance
• VisDoM: Multi-Document QA with Visually Rich Elements Using Multimodal Retrieval-Augmented Generation
• Compressing KV Cache for Long-Context LLM Inference with Inter-Layer Attention Similarity
• Sufficient Context: A New Lens on Retrieval Augmented Generation Systems
• Long²RAG: Evaluating Long-Context & Long-Form Retrieval-Augmented Generation with Key Point Recall
• Should We Really Edit Language Models? On the Evaluation of Edited Language Models
• EPIC: Efficient Position-Independent Caching for Serving Large Language Models
• MIRAGE-Bench: Automatic Multilingual Benchmark Arena for Retrieval-Augmented Generation Systems
• CoFE-RAG: A Comprehensive Full-chain Evaluation Framework for Retrieval-Augmented Generation with Enhanced Data Diversity
• Uncovering Overfitting in Large Language Model Editing
• Assessing Episodic Memory in LLMs with Sequence Order Recall Tasks
• Beyond Prompts: Dynamic Conversational Benchmarking of Large Language Models
• When Context Leads but Parametric Memory Follows in Large Language Models
• MemoRAG: Boosting Long Context Processing with Global Memory-Enhanced Retrieval Augmentation
• RazorAttention: Efficient KV Cache Compression Through Retrieval Heads
• AriGraph: Learning Knowledge Graph World Models with Episodic Memory for LLM Agents
• RankRAG: Unifying Context Ranking with Retrieval-Augmented Generation in LLMs
• BERGEN: A Benchmarking Library for Retrieval-Augmented Generation
• Found in the Middle: Calibrating Positional Attention Bias Improves Long Context Utilization
• Multimodal Needle in a Haystack: Benchmarking Long-Context Capability of Multimodal Large Language Models
• Reversing the Forget-Retain Objectives: An Efficient LLM Unlearning Framework from Logit Difference
• DomainRAG: A Chinese Benchmark for Evaluating Domain-specific Retrieval-Augmented Generation
• Editing the Mind of Giants: An In-Depth Exploration of Pitfalls of Knowledge Editing in Large Language Models
• Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention
• RULER: What's the Real Context Size of Your Long-Context Language Models?
• Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning
• Rebuilding ROME : Resolving Model Collapse during Sequential Model Editing
• Let LLMs Take on the Latest Challenges! A Chinese Dynamic Question Answering Benchmark
• Evaluating Very Long-Term Conversational Memory of LLM Agents
• FeB4RAG: Evaluating Federated Search in the Context of Retrieval Augmented Generation
• Faithfulness vs. Plausibility: On the (Un)Reliability of Explanations from Large Language Models
• LV-Eval: A Balanced Long-Context Benchmark with 5 Length Levels Up to 256K
• CRUD-RAG: A Comprehensive Chinese Benchmark for Retrieval-Augmented Generation of Large Language Models
• CRUD-RAG: A Comprehensive Chinese Benchmark for Retrieval-Augmented Generation of Large Language Models
• MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries
• RECALL: A Benchmark for LLMs Robustness against External Counterfactual Knowledge
• Evoke: Evoking Critical Thinking Abilities in LLMs via Reviewer-Author Prompt Editing
• ReEval: Automatic Hallucination Evaluation for Retrieval-Augmented Large Language Models via Transferable Adversarial Attacks
• Systematic Assessment of Factual Knowledge in Large Language Models
• Towards Verifiable Generation: A Benchmark for Knowledge-aware Language Model Attribution
• Efficient Streaming Language Models with Attention Sinks
• BAMBOO: A Comprehensive Benchmark for Evaluating Long Text Modeling Capacities of Large Language Models
• Evaluating the Ripple Effects of Knowledge Editing in Language Models
• FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning
• Extending Context Window of Large Language Models via Positional Interpolation
• Unifying Large Language Models and Knowledge Graphs: A Roadmap
• Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
• LEACE: Perfect linear concept erasure in closed form
• Enabling Large Language Models to Generate Text with Citations
• FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation
• SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models
• What learning algorithm is in-context learning? Investigations with linear models
• FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
• Locating and Editing Factual Associations in GPT
• Kernelized Concept Erasure
• QAFactEval: Improved QA-Based Factual Consistency Evaluation for Summarization
• Improving language models by retrieving from trillions of tokens
• Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation
• A Dataset for Answering Time-Sensitive Questions
• Deduplicating Training Data Makes Language Models Better
• Time-Aware Language Models as Temporal Knowledge Bases
• The Power of Scale for Parameter-Efficient Prompt Tuning
• BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models
• Dynabench: Rethinking Benchmarking in NLP
• The Pile: An 800GB Dataset of Diverse Text for Language Modeling
• Transformer Feed-Forward Layers Are Key-Value Memories
• Extracting Training Data from Large Language Models
• KILT: a Benchmark for Knowledge Intensive Language Tasks
• Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• On Faithfulness and Factuality in Abstractive Summarization
• ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT
• Don't Stop Pretraining: Adapt Language Models to Domains and Tasks
• Null It Out: Guarding Protected Attributes by Iterative Nullspace\n Projection
• Dense Passage Retrieval for Open-Domain Question Answering
• Asking and Answering Questions to Evaluate the Factual Consistency of Summaries
• REALM: Retrieval-Augmented Language Model Pre-Training
• Machine Unlearning
• Compressive Transformers for Long-Range Sequence Modelling
• Generalization through Memorization: Nearest Neighbor Language Models
• Evaluating the Factual Consistency of Abstractive Text Summarization
• Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
• Language Models as Knowledge Bases?
• Memorizing Normality to Detect Anomaly: Memory-augmented Deep Autoencoder for Unsupervised Anomaly Detection
• Parameter-Efficient Transfer Learning for NLP
• Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context
• The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks
• MS MARCO: A Human Generated MAchine Reading COmprehension Dataset
• Prefix-Tuning: Optimizing Continuous Prompts for Generation
• Prefix-Tuning: Optimizing Continuous Prompts for Generation

Cited by

• Memory for Large Language Models
• Hindsight is 20/20: Building Agent Memory that Retains, Recalls, and Reflects

Related

• Memory for Large Language Models
• ForgetBench: Benchmarking Forgetting Dynamics of Long-Term Parametric Memory in Language Models
• Cognitive Memory in Large Language Models
• Rethinking Memory in LLM based Agents: Representations, Operations, and Emerging Topics
• SCM: Enhancing Large Language Model with Self-Controlled Memory Framework