We would like to measure which pages get read, using PostHog (EU, Frankfurt). It sets a cookie. Nothing is measured unless you agree. See our privacy page.
arXiv sweep · 2025-03 → 2026-08 · built bottom-up from embeddings
Atlas of Machine Memory
Did anyone solve the write path — navigating and remembering beyond the context window without retraining? 7,414 papers, 18 months, one tree.
7,414
papers
248
clusters
5
branches
87
rising
42
fading
SOLVED — and commodity
Semantic lookup. Embedding indexes and hybrid retrieval are settled infrastructure; every memory product converges on the same six engine patterns. Building engines is a graveyard; the value is in owned state.
CONTESTED — no winner
External memory design. 1,471 papers of prosthetics with no benchmark consensus and an open substrate war. Consolidation (what to keep, when to rewrite) is the live fight, not storage.
OPEN — the real seams
Evaluation of memory quality; memory integrity under poisoning; aggregation over corpora that no context can hold; and the native write path itself — test-time training is the only rising candidate inside the weights.
Method note: tree built bottom-up — 7,414 papers embedded locally, clustered (k=248, cohesion 0.86–0.92), each cluster carded by a small model with month histograms, branches and this synthesis written by a larger one. Every claim descends to actual papers.
risingsteadyfadingn = papers in cluster
Making Context Cheaper
2,473 papers · 68 clusters · 27 rising
A third of the entire field optimizes the existing context mechanism: KV-cache compression and quantization, prompt compression, sparse and adaptive attention, and the serving systems that schedule it all. This is the crowded zone — and it is engineering the symptom, not the disease: nothing here changes what context is. Prompt compression reads post-peak (crowded and fading); the live edges are sparse attention learned at pretraining time, task-aware cache compaction that tries to preserve reasoning rather than tokens, and latent-attention variants.
KV Cache Compression Evolutionactive134Prompt & Context Compressioncrowded100Inference Serving & KV Management Systemsactive96Long-Context RL & Scalingcrowded89RAG & Long-Context QA Trade-offsactive83Sparse Attention for Long Inferenceactive80KV-Cache Management for LLM Servingactive80Heterogeneous GPU Serving Infrastructureactive68Long-Context Retrieval & Reasoningactive66KV Cache Compression & Evictionniche65Test-Time Reasoning & Inference Scalingcrowded65Efficient Audio Intelligence & Speech Modelsactive64Gated & Structured Memory for Contextniche62RAG Efficiency & Context Compressionactive60Sparse Attention & KV Cacheactive59Long-Context Evaluation & Engineeringactive58Long-Context Analysis & Diagnosisactive55Dynamic & Adaptive Attentionactive55Diffusion LLM Inference Accelerationcrowded54KV Cache Quantization for Reasoningcrowded53Hierarchical Memory for Long-Video Reasoningactive45Speculative Decoding for Efficient Inferenceniche41Knowledge Editing and Lifelong LLM Updatescrowded40LLM Inference Serving and Schedulingactive39Quantization for Efficient Inferencecrowded38Vision-Language Efficient Compressionactive38Chain-of-Thought Reasoning Compressionactive38On-Device LLM Memory Optimizationactive36Many-Shot Translation & Adaptationactive36Efficient Multi-Head Latent Attentionactive35Language-Specific Encoders & Tokenizationactive35Evidence-Grounded Long-Context QAactive34LLM Inference Parallelism & Efficiencyactive34retrieval, reranking, tool—33Efficient Repository-Level Code Generationactive31Long-Context Evaluation Benchmarkscrowded31Efficient Long-Context Attentionactive30LLM Agent Attack Surfaceactive29Edge-Cloud LLM Serving Infrastructureactive27Efficient Long-Context via Cache Compactionactive26Efficient LLM Training and Inferencecrowded24Jailbreak Detection and Defense for LLMscrowded23Text-to-SQL & Semantic Knowledgeactive22KV Cache & Latent Multi-Agent Communicationactive21Efficient Inference via Pruning & KV-Cachecrowded21Foundation Models & Capabilities Reportscrowded21GPU Inference Performance & System Optimizationcrowded21Efficient Long-Context KV Cachingactive19Distributed Parallelism for Long-Context Trainingcrowded19Positional Encoding & Length Generalizationactive15Dynamic Scheduling for LLM Servingactive11Activation Steering in LLMsactive10Long Video Understandingactive10Long-Form Document Summarizationactive9Mid-Training Paradigm Surveyscrowded9Continual Malware & Code Securityniche9LLM-Assisted Literary Analysisniche8Systems Optimization for LLM Servingactive8Document-Level Machine Translation Contextactive8Efficient Vision-Language Inferenceactive8LLM Serving System Reliabilityactive7Data selection for efficient trainingactive5Legal and professional domain reasoning benchmarksactive5Efficient sentiment and toxicity detectionactive5Long-Context Code Generationactive5Prompt Injection Defenseactive4Carbon-Aware LLM Caching at Cloud Scaleniche3Linguistic Steganography with Anchored Windowsniche1
Prosthetic Memory
1,471 papers · 52 clusters · 23 rising
The write-path attempts: external memory systems bolted onto frozen models. Long-horizon agent memory, conversational-memory benchmarks, self-evolving skill libraries, experience consolidation, temporal knowledge graphs. The tell is that its own benchmark cluster still asks whether ANY unified evaluation exists — dozens of parallel prosthetics, no convergent winner, and the substrate war (vector vs graph vs filesystem memory) openly unresolved. Rising and worth watching: experience consolidation across sessions, memorize-vs-recompute policies, and memory POISONING defense — memory integrity is becoming its own subfield.
Agentic Memory for Long-Horizon Tasksactive73Benchmarking Long-Term Conversational Memoryactive73Long-Term Agent Memory Systemsactive72Self-Evolving Agents & Skill Learningactive71Agentic Multi-Agent Orchestrationcrowded68Hierarchical Planning & Memory-Augmented Agentsactive65Agentic Coding & Multi-Agent Workflowsactive61Self-Evolving Agentic Memory Systemsactive59AGI Architecture & Embodied Cognitionniche53Agent Context & Memory Managementcrowded52Agentic Memory Systemscrowded48Personalized Agent Simulationactive47Neuro-Symbolic Agent Memorycrowded43Graph-Memory GUI Agentsactive42Long-Horizon Agent Memoryactive42Agentic Systems & Tool-Using LLMsactive42Temporal Graphs & Episodic Memoryniche42Agentic RAG for Domain-Specific Reasoningcrowded42memory, agent, agents—33graph, generation, rag—33Agent Memory Poisoning and Defenseactive29Self-Improving Agents via Experienceactive28Self-Improving Agentic Harnessesactive28Agentic Search with Persistent Memoryactive26Medical Agents and Clinical Reasoningactive26Personalized LLM Generation & Memoryactive23Agentic AI & Enterprise Governanceactive22Clinical Reasoning & Adaptive Agentsactive21Multi-Agent Document & Table Analysisactive21Scheduling Multi-Agent LLM Servingcrowded20Agentic Systems Security Attacksactive18Policy Optimization Agent Alignmentactive17Foundation Models & Efficient Scalingcrowded17Long-Horizon Agentic State Reasoningactive14Episodic Memory Multi-Agent Coordinationactive13Episodic Memory in Coding Agentsactive12Context Engineering for Human-AI Collaborationactive11Persistent World Modelsactive11Multi-Agent LLM Communication & Coordinationactive9Affective Memory in Agentic Systemsactive9Vision-Language Navigation Systemsactive6Arbor: Autonomous Search & Navigationactive5Scientific paper to visual media generationactive5Autonomous AI runtime security and self-healingactive5Fleet Provisioning and Autonomous Vehicle Systemsniche4Graph-Based Representation Learningniche2Agentic Planning via Tree Searchniche2Foundation Models for 6G Networksniche2Structured Task Inference Householdsniche1Scene-Aware Visually-Grounded Speechniche1LTL Verification Neural Stateful Agentsniche1Multi-Agent Framework for Related Work Generationniche1
Changing the Machine
1,440 papers · 54 clusters · 17 rising
Architectural bets: linear and hybrid attention, parallelizable recurrence, positional-encoding theory, LoRA-based continual learning, model merging — and test-time training, the one genuinely native write-path idea in the corpus, rising fast. Meanwhile pure linear attention is cooling; its own cards ask why the theory never translated to practice. The hybrid position (mostly-linear with a little real attention) looks like the settling point.
Linear & Hybrid Attention Mechanismsactive90LoRA and Continual Learningactive73Efficient Long-Context Architectures & Recurrenceactive63Continual Learning & Neural Memory Systemsactive61Efficient Transformers & Scalingcrowded57Mixture-of-Experts Serving & Routingcrowded54Parallelizable RNNs & Linear Recurrenceactive54Test-Time Training Adaptationactive52Robust Fine-Tuning & Safety Alignmentcrowded50Streaming Video Generationactive50Multimodal Vision-Language Fusion & Editingactive46Transformer Architecture & In-Context Reasoningactive44Many-Shot In-Context Learningniche42Block Diffusion Language Modelsactive42Vision-Language-Action Robot Learningcrowded41RoPE & Positional Encoding Theoryniche35DNA/RNA Sequence Foundation Modelsactive32State Space Models for Dynamic Graphsactive30Test-Time Adaptation for Vision-Languageactive28Gradient-Based Continual Learning Optimizationniche26Out-of-Distribution Detection Continuallyactive25Robust Multimodal Continual Learningactive24Continual Sparse LLM Trainingactive22Vision-Language Test-Time Adaptationactive22Neural Operators for PDEsniche22Working Memory & Attentionactive22Linear Attention & Delta-Rule Mechanismsactive22Multilingual Low-Resource Continual Learningniche21Long-Horizon Latent Reasoningactive20Spatial Memory for Vision-Language Agentsactive18On-Policy Distillation Long-Contextactive18Multilingual Multimodal Embeddingsactive18KV-Cache Quantization Rotationsniche17Transformer Mean-Field Dynamicsactive17Model Merging & Catastrophic Forgettingactive15Compositional Embeddings & Cognitive Architecturesniche15Hybrid MoE Architectures for Efficient Reasoningcrowded14Real-Time Vision-Language Autonomous Drivingactive14Quantum State-Space Sequence Forecastingniche13Efficient Scaling of Foundation Model Architecturescrowded12Efficient Attention and Distillation for Multimodal Modelscrowded12Active Perception & Visual Memoryactive10Hyperbolic LLMs via Mixture-of-Curvatureactive10Bayesian Optimization & Offline-Online Learningactive9Open-Set Adaptive Object Detectionniche8Corruption as Training Signalactive8Transformer Training Across Domainsniche7Grokking and Generalization Dynamicsactive6Generative Visual Retrieval at Scaleactive6Hybrid Offline-Online Routingactive6LLM Negation & Agent Polarizationniche5State-Space Model Depth Theoryniche4Positional & Semantic Disentanglementactive4Turing Completeness & Finite-Precision Recurrent Netsniche4
The SSM Diaspora
1,111 papers · 44 clusters · 8 rising
Mamba and state-space models in their diffusion phase: the core architecture debate is steady while applications spread into medical imaging, EEG, time series, graphs, and remote sensing — the classic sign of a technology moving from frontier to toolbox. Test-time adaptation lives here too, with its stability-over-long-deployment question still open. Mostly applications; little of it moves our question.
State Space Models & Sequencesactive82Selective State-Space Models & Mambacrowded72Test-Time Adaptationactive67Time Series Forecasting & State-Space Modelsactive66Test-Time Adaptationactive49Vision Mamba Modelsniche48Neural Kalman Filtering & State-Space Learningactive46Mamba for Medical Image Segmentationniche41Mamba & State-Space Graph Modelsniche36Medical AI Agents & Reasoningcrowded36Neural Decoding with State-Space Modelsniche34Urban Spatio-Temporal Forecastingactive34Physics-Informed Energy & Climate Forecastingactive33detection, anomaly, continual—33EEG State-Space Models for Motor Controlactive32Neuromorphic Spiking Networksniche293D Pose & Adaptive Perceptionactive28Mamba-Transformer Hybrid Architecturesactive27State-Space Models for Efficient Inferenceactive23Mamba Geometry & Causal Detectionactive21Time-Series Foundation Models & Forecastingactive21Mamba for Hyperspectral Anomaly Detectionactive20Medical Image Segmentation with Mambaactive20Spatio-Temporal Video Understanding via Memoryactive19Latent Dynamics Under Epistemic Uncertaintyactive19Test-Time Adaptation & Robustnessactive18Wearable Physiological Monitoring via SSMactive17Medical Segmentation Continual Adaptationniche16Time Series Foundation Models Zero-Shotactive16ECG Signal Processing & Cardiovascular AIniche14Deepfake & Multimodal Detection Under Driftniche13AI for Early Cognitive Decline Detectionniche12Medical Image Segmentation & Transfer Learningniche9Robust State-Space Filteringactive8LLM-Driven IoT & System Securityactive8Agentic Trajectory Reasoningniche7GPU Kernel Optimization & Servingactive7Causal Inference & Process Supervisionactive7SAR Multimodal Disaster Assessmentniche6Collaborative Multi-Agent Perceptionactive6Temporal Graph Anomaly Adaptationactive5Transformers for Remote Sensing & Commsniche4Orbital Inference via Solar Computeniche1Privacy-Preserving UAV Swarm Intrusion Detectionniche1
The Old War on Forgetting
904 papers · 25 clusters · 9 rising
Classical continual learning, still grinding: catastrophic-forgetting theory is RISING while replay engineering fades — the signature of a field maturing from recipes toward understanding. Plasticity-stability theory, analytic rehearsal-free methods, federated variants, and a small rising security seam (backdoors that persist through continual updates). The unsolved core is unchanged: cheap, durable, incremental weight updates remain out of reach.
Continual Learning & Parameter Efficiencycrowded100Process-Supervised RL for Agentsactive82Theory of Catastrophic Forgettingcrowded78Multimodal Continual Learning & Replaycrowded70Continual Reinforcement Learning & Adaptationactive60Catastrophic Forgetting & Unlearningactive51Safe Continual Reinforcement Learningactive50Continual Learning & Adaptive Model Mergingactive44Continual Learning with Sleep-Inspired Replaycrowded39Federated Continual Learning Under Driftactive38continual, replay, knowledge—33Plasticity-Stability Tradeoff in Continual Learningactive32Continual Graph Embedding Learningactive29CLIP-Based Incremental Learningniche27Federated Continual Learning Forgettingactive27Knowledge Distillation for Continual Learningactive24Backdoors and Poisoning in Continual Learningniche24Reinforcement Learning Control & Adaptationactive23Analytic Class-Incremental Learningactive21Continual Learning via Adaptive Promptingactive20Quantum Continual Learning Systemsniche8Consciousness-Grounded Continual Learningniche8Preference-Optimized Continual Learningactive8Compositional Open-World Learningniche5Partial Observability & Agent Learning Boundsniche3