AI Memory Research, Explained Simply

TAG routes low-confidence steps to uncertainty-based routing, filters them with guarded acceptance with rollback, chooses between bank selection across rule and exemplar memory, and prunes via evidence-based retirement inside a unified control loop. On SVAMP and ASDiv, TAG reaches 81.0% and 85.2% accuracy, improving over the 74.0% and 77.5% no-memory baselines while a compute-matched Retry baseline stays flat.

Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.

ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.

ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.

AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.

A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.

LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.

Adaptive RAG Memory (ARM) augments a standard retriever–generator stack with a Dynamic Embedding Layer and Remembrance Engine that track usage statistics and apply selective remembrance and decay to embeddings. On a lightweight retrieval benchmark, ARM achieves NDCG@5 ≈ 0.9401 and Recall@5 = 1.000 with 22M parameters, matching larger baselines like gte-small while providing the best efficiency among ultra-efficient models.

Aeon restructures LLM memory using the Atlas, Trace, Semantic Lookaside Buffer, Write Ahead Log, and Sidecar Blob Arena inside a zero copy Core Shell kernel. Aeon achieves 4.70 ns INT8 dot products, 3.09 µs Atlas traversal at 100K nodes, 3.1× compression, and P99 read latency of 750 ns under 16 thread contention compared to FP32 and flat scan baselines.

AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.

Agentic Memory (AgeMem) exposes memory management tools, a three-stage progressive RL strategy, and step-wise GRPO directly inside the agent policy to jointly control long-term and short-term memory. On Qwen3-4B-Instruct, AgeMem attains 54.31% average performance across ALFWorld, SciWorld, PDDL, BabyAI, and HotpotQA, exceeding the best baseline A-Mem at 45.74%.

Agent Memory Below the Prompt stores each agent’s KV state in a block pool, quantizes it via a Q4 pipeline, reloads it with BatchQuantizedKVCache, and reuses it across phases using cross-phase context injection. On Gemma 3 12B, Agent Memory Below the Prompt reduces cold TTFT from 172,096 ms to 1,264 ms at 32K context (136×) compared to FP16 prefix caching baselines like vllm-mlx.

AGENTSYS organizes a Main Agent, Worker Agents, Intent Schemas, and an Alignment Validator into a hierarchical memory system that isolates raw tool outputs and only admits schema-validated JSON. On AgentDojo, AGENTSYS reaches 52.87% attacked utility and 0.78% ASR versus 48.27% and 30.66% for the No Defense baseline.

Neuro-Symbolic Dual Memory Framework uses Progress Memory, Feasibility Memory, a Blueprint Planner Agent, a Progress Monitor Agent, and an Actor Agent to decouple semantic progress guidance from executable feasibility checks. On ALFWorld, Neuro-Symbolic Dual Memory Framework achieves 94.78% success rate versus 88.81% for AWM, and on WebShop reaches 0.7132 score versus 0.5998 for WALL-E 2.0.

All-Mem organizes long-term agent memory through Online/Offline Decoupling, Agentic Topology Consolidation, and Topology-Aware Retrieval over a topology-structured memory bank. On LoCoMo, All-Mem reaches 54.63 4o-J versus Mem0’s 48.91, and on LongMemEval-S All-Mem reaches 60.20 4o-J versus Mem0’s 55.80.

AlpsBench combines Personalized Information Extraction, Personalized Information Update, Personalized Information Retrieval, and Personalized Information Utilization over 2,500 WildChat dialogues with human-verified structured memories. AlpsBench shows, for example, that Gemini-3 Flash scores 51.67 on Task 1 Extraction while DeepSeek Reasoner reaches 0.9569 retrieval recall with 100 distractors on AlpsBench.

AMA orchestrates four agents — the Constructor, Retriever, Judge, and Refresher — to build Raw Text, Fact Knowledge, and Episode Memory and route queries adaptively across these granularities. On the LoCoMo benchmark with GPT-4.1-mini, AMA achieves an overall LLM Score of 0.805 compared to Nemori’s 0.774, while reducing token consumption by approximately 80% relative to FullContext.

A-MBER builds multi-session conversational scenarios via a staged pipeline of persona specification, long-horizon planning, conversation generation, annotation, question construction, and benchmark-unit packaging. On A-MBER, a structured memory system reaches 0.69 judgment accuracy, 0.66 retrieval, and 0.65 explanation versus 0.34, 0.29, and 0.31 for a no-memory baseline.

AMemGym combines Structured Data Generation, On-Policy Interaction, Evaluation Metrics, and Meta-Evaluation to script user state trajectories, drive LLM-simulated role-play, and score write–read–utilization behavior. On AMemGym’s base configuration, AWE-(2,4,30) reaches a 0.291 normalized memory score on interactive evaluation, while native gpt-4.1-mini only achieves 0.203, exposing substantial gaps between memory agents and plain long-context LLMs.

AMV-L manages agent memory using a Memory Value Model, Tiered Lifecycle, Bounded Retrieval Path, and Lifecycle Manager to decouple retention from retrieval eligibility. Under a 70k-request long-running workload, AMV-L improves throughput from 9.027 to 36.977 req/s over TTL and reduces p99 latency from 5398.167 ms to 1233.430 ms while matching LRU’s retrieval quality.

Anatomy of Agentic Memory organizes agentic memory into four structures using components like Lightweight Semantic Memory, Entity-Centric and Personalized Memory, Episodic and Reflective Memory, and Structured and Hierarchical Memory. Anatomy of Agentic Memory then reports comparative results such as Nemori’s 0.781 semantic judge score on LoCoMo versus SimpleMem’s 0.298, and latency differences like 1.129s for Nemori versus 32.372s for MemoryOS.

APEX-EM combines a Procedural Knowledge Graph, Experience Memory store, PRGII workflow, Task Verifiers, and StructuralSignatureExtractor to store and reuse full procedural-episodic traces without changing model weights. On KGQAGen-10k, APEX-EM reaches 89.6% accuracy (95.3% CSR) versus 41.3% without memory and surpasses the GPT-4o w/ SP oracle at 84.9%.

APEX-MEM constructs a semi-structured temporal property graph using Ontology, Entity and Property Resolution, Fact Extraction, and Graph Agents to store and query conversational memory. On LOCOMO, APEX-MEM with GPT5 achieves 88.88% overall accuracy, beating MIRIX at 85.38% by 3.50 percentage points.

ARKV dynamically combines Per-layer OQ ratio estimation, Token importance scoring, and Tri-state cache assignment to manage KV cache precision under a global memory budget. On LongBench, ARKV reaches 0.972 relative performance versus 0.979 for Origin while achieving 4× KV memory reduction and maintaining ~86% Tokens Per Second.

PROACTAGENT combines Experience-Enhanced Online Evolution (EXPONEVO), a structured EXPERIENCE BASE, and Proactive Reinforcement Learning-based Retrieval (PROACTRL) to jointly evolve memory and policy with retrieval as an explicit action. On SciWorld, PROACTAGENT reaches 73.50% SR versus 55.50% for GRPO+Reflexion, while cutting interaction rounds from 27.52 to 18.38.

Mnemonic Sovereignty analyzes long term Write, Store, Retrieve, Execute, Share, and Forget Rollback phases against integrity, confidentiality, availability, and governance objectives for agent memory. Mnemonic Sovereignty’s lifecycle matrix shows most of the ~70 works cluster on write and retrieve integrity, leaving store, availability, and governance primitives like write gate validation and post deletion verification almost entirely unexplored.

ATANT v1.1 structurally analyzes seven benchmarks using the 7 v1.0 continuity properties, the 10 checkpoints, a property-coverage matrix, and the Kenotic v1.0 reference implementation. ATANT v1.1 reports 96% ATANT cumulative-scale versus 8.8% LOCOMO substring accuracy, showing that LOCOMO, LongMemEval, BEAM, MemoryBench, Zep eval, MemGPT/Letta, and RULER measure different properties from continuity.

AtomMem reframes agent memory as a POMDP and composes atomic CRUD operations over a hybrid scratchpad plus vector memory storage using a GRPO-based RL policy. On HotpotQA, 2WikiMultiHopQA, Musique, GAIA, and WebWalkerQA, AtomMem reaches an average 58.8 score, beating MemAgent’s 56.7 with the same Qwen3-8B backbone.

Auxiliary-predicted Compress Memory Model (ApCM Model) combines an Invertible Dimensionality Reduction and Predictor (IDRP) module with a Memory Read-Write Controller, including a global Memory Bank, cosine-similarity read, and access-frequency write policy. ApCM Model achieves lower MSE (0.987171 vs 1.001440) than a Key-Value Memory Network while compressing memory from 1024 to 128 dimensions on random data.

MEMORYARENA orchestrates Memory-Agent-Environment Loops, Multi-Session Working Flow, Bundled Web Shopping, Group Travel Planning, and Progressive Web Search to stress-test how agents store and reuse information across sessions. MEMORYARENA’s main result is that agents with near-saturated scores on long-context benchmarks like LoCoMo still obtain Task Success Rates as low as 0.00–0.12 across its four environments.

BenchPreS combines Contexts, User Profiles, Preference Attributes, Gold Labeling, and an LLM-as-Judge framework to test context-aware preference selectivity in persistent-memory LLMs. BenchPreS shows GPT-5.2 reaches 87.33% Appropriate Application Rate on BenchPreS while still having a 40.95% Misapplication Rate compared to Gemini 3 Pro’s 86.48% Misapplication Rate.

Beyond the Context Window compares Conversation Segmentation, Fact Extraction, Embedding and Storage, and Retrieval Mechanism in a Mem0-based memory system against long-context GPT-5-mini. On LongMemEval, Beyond the Context Window finds LC GPT-5-mini reaches 82.40% accuracy, 33.4 percentage points above the memory system baseline.

Fan Duality Model (FDM) uses the Fan Operator, Local-Global Cache, Freeze-Scan Training, and Holographic Reference Beam Decoding to separate wave-like compression from particle-like associative recall. On WikiText-103, Fan Duality Model (FDM) reaches 64.9 perplexity with Freeze-Scan and 62.79 with holographic decoding, while achieving 0.966 MQAR accuracy compared to Transformer at 0.606.

ByteRover organizes memory through an LLM-driven Agent Layer, a sequential Execution Layer, and a file-based Context Tree with an Adaptive Knowledge Lifecycle. On LoCoMo, ByteRover reaches 96.1% overall accuracy versus 89.9% for HonCho, and on LongMemEval-S ByteRover scores 92.8% overall.

Chronos decomposes dialogue into structured events via the Event Extraction pipeline, stores them in dual event calendar and turn calendar indexes, and uses Dynamic Prompting, Initial Retrieval, and the Chronos Agent for temporal-aware tool-calling. On LongMemEvalS, Chronos Low reaches 92.60% overall accuracy and Chronos High 95.60%, beating EmergenceMem Internal by 7.67 percentage points and Mastra’s OM by 3.02 points.

ClawVM manages agent state as typed pages via the SessionPageTable, RepresentationSelector, FaultObserver, WritebackJournal, and ClawVMEngine inside the agent harness. Across four OpenClaw-derived workloads and six token budgets, ClawVM cuts explicit faults from 67.8 (retrieval baseline) and 1.5 (Compaction-Hybrid) to 0.0 while adding median <50 μs policy-engine overhead per turn.

Codebase-Memory parses repositories with a multi-pass pipeline using the Parse stage, Build stage, Serve stage, FunctionRegistry, Louvain communities, and MCP tool interface to build a persistent SQLite knowledge graph. On a 31-language benchmark, Codebase-Memory reaches 0.83 quality versus 0.92 for an Explorer Agent while using ten times fewer tokens and 2.1 times fewer tool calls.

D-Mem combines Mem0∗, Quality Gating, and Full Deliberation into a dual-process memory system that incrementally stores vector memories and selectively scans raw history. On LoCoMo with GPT-4o-mini, D-Mem’s Quality Gating reaches 53.5 F1 versus the Mem0∗ baseline’s 51.2 F1, recovering 96.7% of the 55.3 F1 Full Deliberation performance with far fewer tokens.

Drawing on Memory uses dual-trace memory encoding, an evidence scoring gate, and a three-state retrieval protocol to store paired fact and scene traces in Letta’s archival memory. On LongMemEval-S, Drawing on Memory reaches 73.7% accuracy versus 53.5% for the fact-only C7-control baseline, a +20.2 percentage point gain concentrated in temporal, update, and multi-session questions.

Experience Compression Spectrum organizes Level 0 Raw Trace, Level 1 Episodic Memory, Level 2 Procedural Skill, and Level 3 Declarative Rule into a unified scaffold-level compression framework. Experience Compression Spectrum’s mapping of 20+ systems and <1% cross-citation rate shows that all existing agents fix a single compression level and never perform adaptive cross-level compression.

GAM builds a Hierarchical Graph Memory Architecture with a global Topic Associative Network, local Event Progression Graphs, State-Based Memory Consolidation, and Graph-Guided Multi-Factor Retrieval to decouple encoding from consolidation. On LoCoMo with Qwen2.5-7B, GAM attains an Average F1 of 40.00 compared to Mem0’s 35.38, and on LongDialQA with Qwen2.5-7B, GAM reaches 12.55 F1 vs MemoryOS at 6.76.

SSGM interposes a Governance Middleware, Read Filtering Gate, Write Validation Gate, and a dual substrate of Mutable Active Graph plus Immutable Episodic Log between agents and memory. SSGM unifies evolving-memory systems into a four-dimensional failure taxonomy and proves that periodic reconciliation can bound semantic drift over infinite horizons.

GradMem combines a WRITE phase, a READ phase, a context encoder Eθ, a self-supervised WRITE objective Lwrite, and a meta-learned initialization M0 to optimize prefix memory tokens via test-time gradient descent while keeping model weights frozen. On associative KV-retrieval with 96 key–value pairs, GradMem with 5 gradient WRITE steps reaches 88.4% exact match versus 12.9% for forward-only RMT with the same 8-vector memory.

HingeMem combines Boundary Guided Long-Term Memory, Dialogue Boundary Extraction, Memory Construction, Query Adaptive Retrieval, Hyperedge Rerank, and Adaptive Stop to segment dialogues into element-indexed hyperedges and plan query-specific retrieval. On LOCOMO, HingeMem achieves 63.9 overall F1 and 75.1 LLM-as-a-Judge score, surpassing the best baseline Zep (56.9 F1) by 7.0 F1 without using category-specific QA formats.

Human-Like Lifelong Memory combines Executive Function and Working Memory, a Memory Service Knowledge Graph, and a Thalamic Gateway to implement dual-process, valence-aware lifelong memory. Human-Like Lifelong Memory is a theoretical framework with seven functional properties and testable predictions rather than benchmark numbers against specific baselines.

Knowledge Packs pre-compute KV Cache Injection, KV–Prefix Equivalence, Banked Routing, and KV Composition to deliver retrieved knowledge and steering via KV states instead of prompt tokens. On HotpotQA, Knowledge Packs’ KV-chat matches RAG at 65.2% EM on Qwen3-8B with 0/500 divergences while eliminating 284 tokens of retrieval text per query.

SleepGate augments transformers with a Conflict-Aware Temporal Tagger, Forgetting Gate, Consolidation Module, and Sleep Trigger that periodically rewrite the KV cache during sleep micro-cycles. On the PI-LLM benchmark, SleepGate achieves 99.5% retrieval accuracy at PI depth 5 and 97.0% at depth 10, while full KV cache, sliding window, H2O, StreamingLLM, and a decay-only ablation all stay below 18% across all depths.

LightMem orchestrates SLM-1 Controller, SLM-2 Selector, SLM-3 Writer, and STM MTM LTM stores to modularize retrieval, writing, and offline consolidation. On LoCoMo, LightMem reaches 34.50 F1 for GPT-4o multi hop questions, +1.64 over A-MEM, while keeping median retrieval latency at 83 ms.

LPC-SM combines local attention, dual-timescale memory, predictive correction, Orthogonal Novelty Transport, and multi-head-coupled residual routing (mHC) inside a single autoregressive block. On OpenWebMath-10k continuation, LPC-SM with adaptive sparse control reaches final LM loss 10.787 versus 12.137 for a fixed sparse controller, a 12.517% improvement.

MAGMA organizes agent memory with an Intent-Aware Router, Adaptive Topological Retrieval, a Data Structure Layer of Relation Graphs and Vector Database, plus dual-stream Synaptic Ingestion and Asynchronous Consolidation. On LoCoMo, MAGMA achieves a 0.700 overall LLM-as-a-Judge score versus 0.590 for Nemori, and reaches 61.2% average accuracy on LongMemEval versus 56.2% for Nemori.

MemEvoBench combines Misleading Memory Injection, Noisy Tool Returns, Biased User Feedback, and a Memory Modification Tool (+ModTool) to stress-test long-term memory safety in LLM agents across 7 domains and 36 risk types. On the QA Style benchmark, MemEvoBench shows Gemini-2.5-Pro’s ASR drops from 67.0% (Vanilla) to 19.0% with +ModTool in Round 1, while biased feedback can push GPT-5’s QA ASR from 59.0% to 78.0% by Round 3.

Memex(RL) optimizes Indexed Experience Memory, CompressExperience, ReadExperience, and ContextStatus so Memex keeps only an indexed summary in-context while archiving full artifacts externally. On modified ALFWorld, Memex(RL) lifts task success from 24.22% to 85.61% over the Memex agent without RL while reducing peak working context from 16,934.46 to 9,634.47 tokens.

MemFactory decomposes memory agents into Module Layer, Agent Layer, Environment Layer, and Trainer Layer with plug and play Extractor, Updater, Retriever, and RecurrentMemoryModule components. On MemAgent eval_50, MemFactory raises Qwen3-1.7B from 0.4727 to 0.5684 and Qwen3-4B-Instruct from 0.6523 to 0.7051 using GRPO.

MemMachine combines Short-term memory, Long-term memory, Profile memory, and the Retrieval Agent to store raw conversational episodes and retrieve clustered context around nucleus matches. On LoCoMo, MemMachine scores 0.9169 with gpt-4.1-mini while using about 80% fewer input tokens than Mem0, and reaches 93.0% on LongMemEvalS with GPT-5-mini.

Memory as Metabolism defines companion knowledge systems with five retention operations (TRIAGE, DECAY, CONTEXTUALIZE, CONSOLIDATE, AUDIT) plus memory gravity and minority-hypothesis retention over a raw buffer, active wiki, and cold memory. Instead of benchmark gains, Memory as Metabolism’s main result is a governance specification that separates descriptive, taxonomic, and normative claims and predicts improved coherence stability, fragility resistance, monoculture resistance, and effective minority-hypothesis influence for companion wikis.

MEMORYCD builds a user memory pool Mu from lifelong Amazon Review histories and evaluates long-context prompting, Mem0, LoCoMo, ReadAgent, MemoryBank, and A-Mem across rating, ranking, and personalized text tasks. On Books and Home & Kitchen, MEMORYCD shows GPT-5 reaches RMSE 0.551–0.624 and NDCG@3 up to 0.610, while Gemini-2.5 Pro peaks at ROUGE-L 0.222 for generation, revealing substantial remaining gaps to real user behavior.

Memory for Autonomous LLM Agents decomposes agent memory into a POMDP-grounded write–manage–read loop, a three-dimensional taxonomy, and five mechanism families spanning context compression, retrieval stores, reflection, hierarchical virtual context, and policy-learned management. Memory for Autonomous LLM Agents synthesizes results like Voyager’s 15.3× tech-tree speedup and MemoryArena’s 80%→45% drop to show that memory architecture often matters more than backbone choice.

Memory in the LLM Era decomposes agent memory into Information Extraction, Memory Management, Memory Storage, and Information Retrieval, then recombines modules into a new hierarchical tree–tier architecture. On LONGMEMEVAL with Qwen2.5-7B, Memory in the LLM Era achieves 38.79 F1 overall versus 36.92 for MemTree.

MemRL structures memory as an Intent-Experience-Utility bank and uses Two-Phase Retrieval plus Runtime Utility Update to learn a value-aware retrieval policy over a frozen LLM. On ALFWorld exploration, MemRL achieves 0.979 success rate compared to 0.921 for MemP, a +0.058 gain with the same frozen GPT-5-mini backbone.

Multi-Agent Memory Architecture organizes Agent IO Layer, Agent Cache Layer, Agent Memory Layer, Agent Cache Sharing, and Agent Memory Access Protocol into a computer-architecture-style design for LLM agents. Multi-Agent Memory Architecture’s main result is a conceptual unification of shared and distributed memory plus a research agenda for multi-agent memory consistency instead of benchmark gains.

OS-SYMPHONY coordinates an Orchestrator, Reflection-Memory Agent, and Versatile Tool Agents (Multimodal Searcher, Grounders, Coder) to stabilize long-horizon GUI workflows and fetch visual tutorials on demand. On OSWorld-Verified, OS-SYMPHONY with GPT-5 scores 65.84% at 100 steps, beating Agent S3 w/ GPT-5 (62.63%) by 3.21 percentage points.

Self-Correcting RAG combines a MMKP-based Context Selector, NLI-Guided MCTS Generator, and Self-Correcting RAG Optimizer to jointly optimize retrieval and reasoning under token and redundancy constraints. On six QA datasets, Self-Correcting RAG attains average EM 37.1 and F1 45.8, beating the strongest baseline CRAG (avg EM 34.3, F1 43.3).

SE-Search combines Memory Purification, Atomic Query, and Dense Rewards inside a GRPO-trained search agent that follows a Think-Search-Memorize-Answer loop. On seven QA benchmarks, SE-Search-3B attains 0.420 average EM, a 10.8 point absolute and 33.8% relative gain over Search-R1-Base.

ACON combines History Compression, Observation Compression, Compression Guideline Optimization, and Compressor Distillation to rewrite agent histories and observations into concise, task-aware summaries. On AppWorld, ACON UTCO with gpt-4.1 achieves 56.5% accuracy with 7.33k peak tokens, versus 56.0% accuracy with 9.93k peak tokens for No compression.

A-MEM organizes agent memory via Note Construction, Link Generation, Memory Evolution, and Retrieve Relative Memory to build an evolving, interconnected note graph. On the LoCoMo dataset, A-MEM with GPT-4o-mini reaches 27.02 F1 on Multi Hop questions, +17.87 over ReadAgent, while cutting average token length from 16,910 to 2,520.

A-MemGuard combines consensus-based validation, dual-memory structure, lesson memory, and path divergence scoring to sanitize retrieved memories and revise actions using past failures. On EHRAgent under AgentPoison, A-MemGuard reduces ASR-r from 100.0% to 2.13% and ASR-t from 100.0% to 6.38%, far below LLM Auditor and Distil Classifier.

LIGHT augments LLMs with Retrieval from the Conversation, Scratchpad Formation and Utilization, and a Working Memory buffer plus noise filtering to answer BEAM’s long-context probing questions. On the BEAM benchmark, LIGHT raises GPT-4.1-nano’s average score at 10M-token conversations from 0.109 to 0.226, a +107.3% gain over the vanilla long-context baseline.

CAM builds hierarchical schemata using an incremental overlapping clustering algorithm, ego centric disentanglement, and a Prune and Grow associative strategy for retrieval. On NovelQA, CAM achieves 52.3 ACC-L versus RAPTOR’s 47.8, a +4.5 point gain while also improving efficiency in long-text reading comprehension.

Evaluating Long-Term Memory for Long-Context Question Answering compares Full Context, RAG, A-Mem, RAG+PromptOpt, and RAG+EpMem memory components across semantic, episodic, and procedural memory for long conversational QA. On LoCoMo, RAG+EpMem reaches an average F1 ranking of 1.83 for Llama 3.2-3B Instruct and 1.80 for GPT-4o mini while using around 1,000 tokens per query versus over 23,000 for Full Context.

MemoryAgentBench standardizes multi-turn datasets into chunked conversations with memorization prompts, then evaluates long-context agents, RAG agents, and agentic memory agents across Accurate Retrieval, Test-Time Learning, Long-Range Understanding, and Selective Forgetting. On the overall score in Table 3, the GPT-4.1-mini long-context agent reaches 71.8 on Accurate Retrieval tasks compared to 49.2 for the GPT-4o-mini long-context baseline.

MaRS organizes agent memory into episodic, semantic, social, and task nodes with provenance, scored by a privacy-aware retention controller and governed by FIFO, LRU, Priority Decay, Reflection-Summary, Random-Drop, and Hybrid policies. On the FiFA benchmark, the Hybrid policy in MaRS achieves a composite score of ≈0.911 across 300 runs and five memory budgets, outperforming simpler policies while preserving privacy and cost efficiency.

From Human Memory to AI Memory maps human memory categories onto AI memory using the 3D-8Q taxonomy with Personal Memory, System Memory, and the Three-Dimensional Eight-Quadrant Memory Taxonomy. The main result is that From Human Memory to AI Memory systematically organizes memory in LLM-driven AI systems across eight quadrants defined by object, form, and time, connecting them to human memory types.

General Agentic Memory (GAM) combines a Memorizer, Researcher, page-store, and memory to keep full trajectories while constructing lightweight guidance for deep research. On RULER 128K retrieval, GAM achieves 97.70% accuracy compared to 94.25% for RAG using GPT-4o-mini, while also reaching 64.07 F1 on HotpotQA-56K.

HINDSIGHT organizes agent memory into four networks via TEMPR and layers CARA on top to retain, recall, and reflect with explicit opinions and behavioral profiles. On LongMemEval, HINDSIGHT with Gemini-3 Pro scores 91.4% overall versus 60.2% for full-context GPT-4o, while HINDSIGHT with OSS-20B jumps from 39.0% to 83.6% over a full-context OSS-20B baseline.

IMDMR combines a Memory Storage Layer, Multi-Dimensional Search Engine, Intelligent Query Processor, and Response Generation Module to retrieve conversational memories across semantic, entity, category, intent, context, and temporal dimensions. On the synthetic 1,000 conversation benchmark, IMDMR-Prod achieves an overall score of 0.792 compared to 0.207 for spaCy + RAG, a 3.8x improvement.

HGMEM represents working memory as a hypergraph with Hypergraph-based Memory Storage, Adaptive Memory-based Evidence Retrieval, and Dynamic Memory Evolving to build high-order correlations across entities and facts. On Prelude long narrative understanding, HGMEM with GPT-4o achieves 73.81% accuracy compared to 72.22% for HippoRAG v2, while also reaching 69.74 comprehensiveness on Longbench generative sense-making QA.

Reflective Memory Management (RMM) uses a memory bank, retriever, reranker, and LLM to implement Prospective Reflection and Retrospective Reflection for topic-based storage and RL-based retrieval refinement. On LongMemEval, RMM with GTE achieves 69.8% Recall@5 and 70.4% accuracy, compared to 62.4% Recall@5 and 63.6% accuracy for GTE RAG.

Latent learning uses oracle episodic retrieval, within-experience in-context learning, parametric learning, and latent learning benchmarks to study how reinstating full past episodes into context changes generalization. Latent learning shows that oracle retrieval solves latent tests like reversals and codebooks where parametric-only transformers stay near 0% despite strong performance on forward and in-context variants.

Learning from Supervision with Semantic and Episodic Memory combines a performance agent, critic agent, semantic memory, episodic memory, and memory retriever to turn label-grounded critiques into reusable supervision without parameter updates. On the Multi-Condition Ranking dataset with Mixtral 8x22B and o4-mini as critic, Learning from Supervision with Semantic and Episodic Memory reaches 85.6% accuracy, a 24.8% gain over the EP_LABEL baseline at 60.8%.

LiCoMemory organizes long term dialogue with CogniGraph, Query Processing and Integrated Rerank, and Real Time Interactions to keep session summaries, triples, and chunks linked. On LongMemEval with GPT-4o-mini, LiCoMemory reaches 73.80% accuracy and 76.63% recall, beating Mem0g by 9.0 and 7.1 points.

LightMem pipelines a Cognitive-Inspired Sensory Memory, Topic Segmentation Submodule, Topic-Aware Short-Term Memory, and Long-Term Memory with Sleep-Time Update to filter, group, summarize, and asynchronously consolidate dialogue history. On LongMemEval-S with Qwen3-30B-A3B-Instruct-2507, LightMem reaches 70.20% ACC vs 65.20% for A-MEM (+5.00 points) while reducing total token usage by up to 21.8× and API calls by up to 17.1×.

LLM-Empowered Embodied Agent for Memory-Augmented Task Planning in Household Robotics coordinates a routing agent, task planning agent, and knowledge base agent over RAG and ChromaDB to translate household commands into grounded robot actions. In three tabletop scenarios, Qwen2.5-32B in LLM-Empowered Embodied Agent for Memory-Augmented Task Planning in Household Robotics achieves 84.3% total lenient task planning accuracy versus 68.7% for Gemma2-27B and 61.1% for LLaMa3.1-8B.

MEM1 trains an RL agent that updates a compact Internal State (IS) using Masked Trajectory for Policy Optimization, 2D Attention Mask, and Multi-Objective Task Design to jointly consolidate memory and reason. On 16-objective multi-hop QA, MEM1-7B reaches EM 1.97 vs 0.567 for Qwen2.5-14B-Instruct while using 10.4×10² peak tokens vs 38.4×10².

MemBench evaluates LLM-based agents with multi-scenario datasets, multi-level memory content, and a time-aware benchmark using components like Multi-scenario Dataset, Multi-level Memory Content, and Multi-metric Evaluation. MemBench shows that mechanisms such as GenerativeAgent, MemGPT, MemoryBank, and SCMemory can drop from accuracies around 0.7 on 10k-token settings to roughly 0.3–0.4 at 100k tokens, exposing clear capacity limits.

Memento combines a Planner, Executor, Case Memory, Subtask Memory, and Tool Memory inside a memory-based MDP, learning a neural case-selection policy over a growing Case Bank instead of updating LLM weights. On GAIA, Memento achieves 87.88% Pass@3 on validation and 79.40% on the test set, surpassing DeepResearcher’s 51.8% F1 and 60.5% PM on DeepResearcher benchmarks by +14.8 F1 and +19.9 PM.

MemEvolve decomposes agent memory into Encode, Store, Retrieve, and Manage modules and meta evolves these components via a dual evolution process over candidate architectures. On xBench DeepSearch, MemEvolve with GPT 5 mini raises Flash Searcher pass@1 from 69.0 to 74.0 and WebWalkerQA accuracy from 58.82 to 61.18 while keeping API cost near 0.141 per query.

MemInsight augments agent memory using Attribute Mining, Annotation and Attribute Prioritization, and Memory Retrieval modules that generate and exploit structured attributes over past interactions. On the LoCoMo question answering benchmark, MemInsight with Claude-3-Sonnet priority augmentation achieves 60.5% Recall@5 versus 26.5% for DPR, a 34.0-point improvement.

Memoria augments LLM chats with structured conversation logging, dynamic user persona via KG, session level memory for real time context, and seamless retrieval for context aware responses to provide persistent, interpretable memory. On LongMemEvals single-session-user and knowledge-update subsets, Memoria reaches 87.1% and 80.8% accuracy respectively, surpassing A-Mem (OpenAI) while using much shorter prompts.

MemoriesDB stores each Memory Record, Edges and Relations, and the Temporal Semantic Stack inside PostgreSQL with pgvector, exposing unified temporal–semantic–relational queries. MemoriesDB’s main result is a working implementation that demonstrates scalable time-bounded recall and hybrid semantic–structural queries on commodity SQL infrastructure without specialized vector or graph engines.

Memorization to Generalization recasts diffusion training and sampling as Dense Associative Memory dynamics, analyzing memorized, spurious, and generalized states via energy basins and curvature. Memorization to Generalization shows that as training size grows on MNIST, CIFAR10, FASHION-MNIST, LSUN-CHURCH, and Stable Diffusion, spurious states peak at the memorization–generalization boundary and have distinct basin volume and curvature signatures.

Phonetic Trajectory Memory (PTM) combines the Acoustic Injection, Entropy Filter, Neuro-Symbolic Relay, and Resonance Engine to encode text as a continuous trajectory on an ergodic Hyper-Torus Memory instead of a growing KV cache. PTM delivers >3,000× signal-to-KV compression while maintaining ≈92% factual accuracy and sub-50 ms retrieval latency on long narrative and scientific corpora compared to dense KV baselines.

DM-RAG augments Phi-4-mini with a Short-Term Memory (STM) buffer, Long-Term Memory (LTM) FAISS store, Bayesian fusion, and a logistic regression confidence model for structured log analysis. On UNSW-NB15, DM-RAG reaches 98.70% recall and 69.59% F1, beating the Phi-4 + RAG (MITRE) baseline in F1 by 17.89 points.

Memory-Augmented Transformers organizes functional objectives, memory types, and integration techniques into a unified taxonomy that connects biological memory principles with concrete architectures like Memformer, Titans, ATLAS, and EMAT. Memory-Augmented Transformers’ main result is a systematic three-dimensional classification that links dynamic multi-timescale memory, selective attention, and consolidation to specific Transformer designs and emerging lifelong-learning paradigms.

MemoryBench orchestrates Task Provider, User Simulator, and Performance Monitor to feed heterogeneous tasks, simulate explicit and implicit feedback, and score LLM systems across declarative and procedural memory. MemoryBench’s main finding is that state-of-the-art memory systems like A-Mem, Mem0, and MemoryOS often fail to beat naive BM25 or embedding-based RAG on partitions such as SiLo and LiLo.

Memory Decoder combines a Pre-training stage that aligns with kNN-LM distributions and an Inference interpolation mechanism that mixes Memory Decoder and base LLM outputs without changing base parameters. On Wikitext-103, Memory Decoder with 124M parameters reaches 13.36 perplexity on GPT2-small versus 14.76 for DAPT, and on specialized domains a single 0.5B Memory Decoder reduces average perplexity from 14.88 to 4.05 on Qwen2-0.5B.

MemVid combines a memory model, memorizer, retriever, and generator so a reasoning-oriented KV-cache memory can produce task-specific clues that drive video moment retrieval. On VideoMME (with subtitles, Avg column), MemVid scores 65.7% compared to 61.0% for Video-XL-7B, a +4.7 point gain under the same 7B scale.

Memory in the Age of AI Agents formalizes agent memory with Memory Formation, Memory Evolution, and Memory Retrieval operators, and classifies memories into token-level, parametric, and latent forms plus factual, experiential, and working functions. Memory in the Age of AI Agents’ main result is a unified Forms–Functions–Dynamics framework that consolidates fragmented LLM agent memory work, benchmarks, and open-source frameworks into a coherent taxonomy.

MemOS standardizes memory as MemCube units orchestrated by MemReader, MemScheduler, MemLifecycle, MemOperator, MemVault, and MemGovernance to manage parametric, activation, and plaintext memory as one system. MemOS does not report benchmark numbers but instead contributes a unified architecture and Memory Interchange Protocol for cross-LLM memory sharing and lifecycle governance.

Memp constructs agent skills via Build, Retrieve, and Update modules that turn past trajectories into scripts, trajectories, and combined proceduralizations stored in a procedural memory library. On ALFWorld, Memp’s proceduralization with GPT-4o reaches 77.86% test success versus 42.14% with no memory, while reducing steps from 23.76 to 15.01.

MEPIC extends vLLM with a Chunk Cache Coordinator, Chunk Matcher, Hybrid KV Manager, Chunk LRU Manager, and Chunk Processor to manage canonical, page-aligned, position-independent KV chunks in HBM. On long-context workloads, MEPIC reduces HBM usage by up to 5.21× and lowers latency by up to 11.48% compared to CacheBlend on Mistral-7B-Instruct-v0.3.

MeVe decomposes retrieval into Initial Retrieval, Relevance Verification, Fallback Retrieval, Context Prioritization, and Token Budgeting to tightly control what enters the LLM context. On a Wikipedia subset and HotpotQA, MeVe reduces average context from 188.8 to 79.8 tokens and from 308.6 to 78.5 tokens respectively compared to Standard RAG while keeping retrieval time comparable.

M+ augments short-term memory θ, long-term memory Θ, a co-trained retriever, and a Multi-LoRA design on top of MemoryLLM’s layer-wise memory pools. On SQuAD-style knowledge retention, M+ maintains accuracy beyond 160k tokens while MemoryLLM-7B collapses before 20k and Llama-3.1-8B-SnapKV fails beyond 30k tokens.

MIRIX organizes Core Memory, Episodic Memory, Semantic Memory, Procedural Memory, Resource Memory, and Knowledge Vault under multi-agent control with Active Retrieval for topic-driven access. On ScreenshotVQA, MIRIX reaches 0.5950 accuracy vs 0.4410 for SigLIP@50 while shrinking storage from 15.07GB to 15.89MB, and on LOCOMO MIRIX scores 85.38% vs 79.09% for Zep.

MMAG organizes conversational memory, long-term user memory, episodic and event-linked memories, sensory and context-aware memory, and short-term working memory under a modular memory controller integrated with Heero’s encrypted Firestore and S3 stores. MMAG delivers a 20% increase in user retention and a 30% increase in average conversation duration on the Heero language learning platform compared to its pre-memory deployment.

Mnemosyne combines a Commitment pipeline with substance and redundancy filters, a probabilistic Recall traversal over a graph-structured store, asynchronous Core Summary updates, and a Pruning module to manage long-term memory on edge devices. On the LoCoMo benchmark, Mnemosyne reaches 60.42% temporal reasoning J-score and a 54.55% overall J-score, compared to 51.55% temporal reasoning and 62.74% overall for Memory-R1, and achieves a 65.8% win rate over a 31.07% naive RAG baseline in human evaluations.

MobileRAG builds a fully on-device RAG stack by integrating EcoVector, Selective Content Reduction, DB Construction, and Chat Application over a local SQLite store. On SQuAD with Qwen2.5 1.5B, MobileRAG reaches 65.1% accuracy while reducing TTFT from 11.78s for Advanced RAG to 7.41s and cutting power from 89.09J to 56.28J.

MR.Rec unifies User-specific Local Memory, Cross-user Global Memory, Reasoning-enhanced Memory Retrieval, and Reinforcement Learning for Memory-synergized Reasoning into a single recommendation assistant pipeline. On the Amazon-C4–based benchmark, MR.Rec achieves NDCG@100 = 0.113 and Recall@100 = 0.270, improving over the best baseline Rec-R1 (NDCG@100 = 0.104, Recall@100 = 0.260).

Muon Outperforms Adam in Tail-End Associative Memory Learning analyzes how VO attention weights, FFN matrices, normalized SVD entropy, and effective rank behave under Muon versus Adam in transformer associative memories. Muon Outperforms Adam in Tail-End Associative Memory Learning finds that applying Muon to VO and FFN nearly recovers full-Muon validation loss (3.5654 vs 3.9242 for All Adam at 10k steps on FineWeb) while improving tail-class accuracy on a heavy-tailed QA task compared to Adam and SGD+Momentum.

PersonaMem-v2 combines PERSONAMEM-V2: IMPLICIT PERSONAS, RL with Long-Context Reasoning, RL with Agentic Memory, and a User Privacy-Aware Design to train Qwen3-4B with GRPO on implicit user preferences from long, noisy histories. PersonaMem-v2 achieves 55.2% MCQ and 60.7% open-ended accuracy on PERSONAMEM-V2, surpassing GPT-5-Chat’s 45.6% and 46.2% while using a 2k-token agentic memory instead of full 32k–128k contexts.

Episodic Memory is the Missing Piece for Long-Term LLM Agents proposes an architecture where in-context memory, external memory, and parametric memory are coordinated via consolidation, encoding, and retrieval to realize long-term, instance-specific, contextual episodic traces. Episodic Memory is the Missing Piece for Long-Term LLM Agents contributes a five-property taxonomy, a three-way memory categorization (in-context, external, parametric), and a roadmap of six research questions instead of benchmark gains.

PREMem builds long term dialogue memory by combining Episodic Memory Extraction, Pre Storage Memory Reasoning, semantic clustering, a persistent memory pool, and an inference phase over enriched memory fragments. PREMem reaches 71.4 LLM as a judge on LongMemEval with gpt 4.1 base, a +15.5 gain over HippoRAG 2 and +9.6 over A Mem.

ReMe manages procedural memory through experience acquisition, experience reuse, and experience refinement, combining multi-faceted distillation, context-adaptive reuse, and utility-based deletion into a single lifecycle. On BFCL-V3 and AppWorld, Qwen3-8B with ReMe (dynamic) achieves 34.94% Avg@4 vs 27.65% for the No Memory baseline, and 55.03% Pass@4 vs 46.20%, showing that self-evolving memory can substitute for model scale.

RGMem builds a multi-scale memory state using Microscopic Evidence Space DL0, Structured Knowledge Space G, and renormalization operators RK1, RK2, RK3 to evolve user profiles. On PersonaMem with GPT-4.1, RGMem reaches 74.01% Avg., beating Memory OS by 8.98 points.

Semantic Anchoring enriches conversational memory by combining a hybrid memory store with dense and symbolic indexes, structured memory representation tuples, hybrid storage and indexing, and a retrieval scoring method. On MultiWOZ-Long, Semantic Anchoring reaches 83.5% Factual Recall and 80.8% Discourse Coherence, beating Entity-RAG by 7.6 and 8.6 points respectively.

SGMem organizes long conversations via SGMem Construction and Management, SGMem Usage, sentence level graphs, and multi hop retrieval over sessions, rounds, turns, summaries, facts, and insights. SGMem achieves 0.700 Accuracy (Top 5) on LongMemEval and 0.526 on LoCoMo, beating the RAG-SMFI baseline at 0.676 and 0.510 respectively.

TeleMem converts interactions into unified semantic nodes via the representation layer, organizes them in a memory graph with Insert and ReInsert, and reads them using closure-based retrieval and a ReAct-style multimodal agent. On ZH-4O, TeleMem reaches 86.33% QA Accuracy, beating the Mem0 baseline at 70.20% and the RAG baseline at 62.45%.

Test-time regression uses memorization as regression, memory retrieval, and test-time regression layers to reinterpret sequence architectures as solving a regression problem over key value pairs during the forward pass. This unification shows how linear attention, state space models, fast weight programmers, online learning layers, and softmax attention are all instances of the same framework and explains phenomena like linear attention’s failures and the role of query key normalization.

TurboQuant combines MSE Optimal TurboQuant, Inner-product Optimal TurboQuant, QJL, Random Rotation Matrix Π, and Lloyd-Max Quantizer to quantize vectors online with near-optimal distortion-rate guarantees. TurboQuant matches the Shannon lower bound within a factor of √(3π/2)≈2.7 for MSE and achieves absolute quality neutrality for KV cache quantization at 3.5 bits per channel compared to full-precision baselines.

PI-LLM uses a synthetic key–value retrieval task, Interference Endurance Score (IES), per-key forget prompts, and a mock QA reset to stress-test working-memory-like behavior under proactive interference. PI-LLM finds a universal log-linear decay in retrieval accuracy across 0.6B–637B-parameter LLMs as interference grows, revealing that parameter size, not context window length, predicts interference robustness.

Understanding Transformer from the Perspective of Associative Memory reframes Softmax Attention, Linear Attention, FFN, and DeltaNet as instances of a unified associative memory with explicit memory capacity and update rules. Using this lens, Understanding Transformer from the Perspective of Associative Memory derives retrieval SNR for different kernels, unifies attention and FFNs, and proves that DeltaFormer achieves circuit complexity beyond TC0, reaching NC1 expressivity.

Understanding Users' Privacy Perceptions Towards LLM's RAG-based Memory analyzes users' mental models, privacy calculus, and expectations around RAG-based memory across generation, management, usage, and updating. Understanding Users' Privacy Perceptions Towards LLM's RAG-based Memory finds users demand explicit consent, fine-grained editing and deletion, and visibility into inferred information to trust RAG-based memory systems.

MEXTRA crafts black box attacking prompts and automated diverse prompt generators that target the memory module, similarity scoring function, retrieval depth, memory size, and LLM backbone. MEXTRA extracts 50 queries from a 200 record EHRAgent memory and 26 from RAP, with extracted efficiency up to 0.42 compared to weaker baselines without workflow aligned prompts.

WebATLAS combines a Planner, Actor, Critic, and Multi-layered Memory (Working Memory, Cognitive Map, Semantic Memory) to simulate and score actions before executing them on the web. On WebArena-Lite, WebATLAS achieves 63.0% average success versus 53.9% for Plan-and-Act, a +9.1 point gain without website-specific fine-tuning.

WorldMM dynamically coordinates Episodic Memory, Semantic Memory, Visual Memory, an Adaptive Retrieval Agent, and a Response Agent to answer queries over hour- to week-long videos. On five long video QA benchmarks, WorldMM-GPT reaches 69.5% average accuracy, beating M3-Agent’s 55.1% by 14.4 points and the best prior memory baseline HippoRAG’s 57.0% by 12.5 points.

Zep builds memory using Episode Subgraph, Semantic Entity Subgraph, Community Subgraph, and a three-stage Search–Reranker–Constructor pipeline over the Graphiti temporal knowledge graph. On LongMemEval, Zep with gpt-4o scores 71.2% vs a 60.2% full-context baseline and reduces average latency from 28.9 s to 2.58 s.

AriGraph builds a joint world model by combining semantic memory, episodic memory, semantic search, episodic search, and the Ariadne cognitive architecture into a single evolving graph. On NetHack, AriGraph lets Ariadne reach a score of 593.00 with room-only observations, compared to 341.67 for NetPlay with the same input and 675.33 for NetPlay with full level observations.

EMG-RAG combines Editable Memory Graphs, a two-stage MDP for Selecting Memories on EMGs, and Data Collection from real smartphone assistants to support editable, graph-structured personal memories. On a 2,500-user business dataset, EMG-RAG with GPT-4 reaches 75.99 BLEU on question answering, a +11.83 gain over M-RAG.

Deciphering the Interplay of Parametric and Non-parametric Memory instruments causal mediation analysis, Experiment 1, Experiment 2, and Path Specific Effects (PSE) inside ATLAS to trace how parametric and non-parametric memories compete token-by-token. Deciphering the Interplay of Parametric and Non-parametric Memory reports a strong shift toward counterfactual answers in altered contexts, with a t-test p-value of 1.60e-4 and Cohen’s d of -0.9851 for non-parametric versus parametric behavior.

Do LLMs dream of elephants studies how a self-attention layer, value matrix, embedding matrix, latent concept association task, and context hijacking prompts interact to implement associative memory in transformers. Do LLMs dream of elephants proves theoretically (Theorem 1, Theorem 4) and empirically that a one-layer transformer can achieve arbitrarily small error on latent concept association by using the value matrix as associative memory.

Dual Memory Networks combines a Dynamic Memory Network, Static Memory Network, a shared ReadOut module, Projection Layers ω, and a Memory Interactive Strategy to build sample-adaptive classifiers on top of frozen CLIP encoders. On zero-shot ImageNet with ViT-B/16, Dual Memory Networks achieves 72.25% accuracy vs 66.73% for CLIP and 68.98% for TPT, a +5.52 and +3.27 point gain respectively.

DynaMem maintains a Dynamic 3D Voxel Map, supports Embedded Vision Language Features and Multimodal Large Language Models querying, and exposes Exploration Primitives and an Obstacle map for navigation and manipulation. On real Stretch SE3 experiments, DynaMem achieves a 70% success rate on dynamic pick-and-drop tasks compared to 30% for the static OK-Robot baseline.

EMU combines a semantic memory embedding via deterministic conditional autoencoder and an episodic incentive built from desirability in the episodic buffer to guide cooperative MARL. EMU improves learning speed and final win-rates on StarCraft II SMAC and Google Research Football compared to QPLEX, CDS, and EMC, especially on super hard maps.

Embodied-RAG builds a multimodal Topological Map and a hierarchical Semantic Forest and then runs Top-down Retrieval with LLM-based selection and hybrid re-ranking to drive Generation of waypoints and explanations. On the E-multimodal Embodied-Experiences dataset, Embodied-RAG reaches P(Q|A)=0.67 for implicit queries (Q only), compared to 0.13 for LightRAG, while building graph memory 9.76× faster than LightRAG.

MemoAnalyzer analyzes past inputs and long-term memories using prompt-based privacy inference, confidence and sensitivity visualization, and source tracking with an editing proxy. In a 5-day study on work, life, and academic tasks, MemoAnalyzer reduced total inferred private information by 22.3% compared to GPT memory while keeping completion time comparable to GPT and Manual baselines.

HiAgent manages working memory using Subgoal based Hierarchical Working Memory, Observation Summarization, and a Trajectory Retrieval module to chunk and selectively expand action observation histories. On five AgentBoard long horizon tasks, HiAgent reaches 42.00% overall success rate versus 21.00% for STANDARD, with 3.80 fewer average steps.

HippoRAG combines Offline Indexing, Online Retrieval, a schemaless knowledge graph, retrieval encoders, and node specificity to mimic hippocampal indexing for long-term memory. On 2WikiMultiHopQA, HippoRAG with Contriever achieves 89.5 R@5 vs 68.2 for ColBERTv2, a +21.3 point gain in single-step retrieval.

EM-LLM organises long token streams into episodic events via surprise-based segmentation, boundary refinement, and a two-stage memory retrieval with similarity and contiguity buffers. On LongBench with LLaMA-3.1-8B, EM-LLM reaches 51.58% average score versus 39.3% for full-context processing and 36.44% for RAG.

Ewe augments Llama-3.1 with an explicit working memory, real-time feedback, fact-checking outcomes, and relevant knowledge memories that are refreshed via FIFO KV-cache updates during decoding. On the Biography dataset, Ewe reaches 49.7 VeriScore F1 versus 37.1 for Llama-3.1 70B (+12.6), while keeping AlpacaEval win rate around 50% against the same baseline.

Larimar couples a BERT large encoder, a deterministic hierarchical memory module, a scope detector, and a GPT2-large or GPTJ-6B decoder via learned projection WM to perform memory-conditioned generation. On CounterFact and ZsRE, Larimar attains up to 100.0% edit success and 0.97 edit retention rate while being 4–10× faster than ROME and GRACE.

LongMemEval evaluates long-term interactive memory by running chat assistants through indexing, retrieval, and reading over 50k sessions with fact-augmented keys and time-aware query expansion. On LONGMEMEVALS, long-context LLMs like GPT-4o, Llama 3.1, and Phi-3 suffer 30%–60% accuracy drops compared to oracle evidence-only reading, revealing severe limitations in current long-context designs.

Long Term Memory equips OMNE with a Data Framework for LTM, Development Framework for LTM, and multi-agent collaboration over personalized memories to support AI self-evolution. On the GAIA benchmark, OMNE reaches first place, showing that LTM-driven multi-agent personalization can solve complex real-world tasks better than prior agent systems.

MemLong augments OpenLLaMA-3B with a Ret-Mem module, Memory Bank, Retriever, Retrieval Causal Attention, and Dynamic Memory Update to store and retrieve chunk-level K-V caches via dense embeddings. On PG19 at 16k tokens, MemLong with 32K Memory reaches perplexity 9.73 vs 10.37 for MemLong-3B* without memory, and achieves up to +10.2 percentage points over OpenLLaMA on retrieval-augmented in-context learning.

MemoNav composes Short-term memory, Selective forgetting module, Long-term memory, Working memory generation, and Transformer decoders to focus navigation on goal-relevant topological map nodes. On Gibson 1-goal, MemoNav reaches 74.7% SR vs 70.0% for VGM, and on Gibson 4-goal multi-goal tasks MemoNav achieves 28.9% PR vs 21.5% for VGM.

PerLTQA builds a synthetic personal memory database and runs questions through Memory Classification, Memory Retrieval, and Memory Synthesis to test how LLMs use semantic and episodic memories. On the PerLTQA benchmark, BERT-base achieves 95.7 F1 for memory classification, while gpt-3.5-turbo reaches MAP 0.756 for memory synthesis with retrieval and classification.

MD-DETR augments Deformable-DETR with frozen encoder–decoder, task-partitioned memory modules, localized query retrieval, and background thresholding to mitigate background relegation. On MS-COCO continual detection, MD-DETR reaches 50.2 mAP@A on Task 4, +11.0 over CL-DETR and +14.8 over ERD.

Retrieval-Augmented Decision Transformer (RA-DT) combines a vector index, embedding model g(·), maximum inner product search, experience reweighting, and cross-attention layers to retrieve and fuse relevant sub-trajectories into a Decision Transformer policy. On Dark-Room 10×10, RA-DT reaches near-optimal average reward over 40 in-context trials while using a 50-step context window, whereas baselines like Algorithm Distillation require entire episodes of up to 100 steps.

SAGE coordinates Iterative Feedback, Reflection, Short-Term Memory, Long-Term Memory, and MemorySyntax so the assistant, checker, and user co-evolve policies and memories over time. On AgentBench and long-context QA like HotpotQA, SAGE lifts GPT-3.5’s Database score from 25.9 to 37.6 and HotpotQA answer accuracy from 48.5% to 68.3%.

WM-Neurosymbolic combines External Working Memory, Working Memory Initialization, Symbolic Rule Grounding, and LLM-based Rule Implementation to execute multi-step rule application with explicit fact and rule tracking. On CLUTRR, WM-Neurosymbolic reaches 92.34% accuracy with GPT-4, a +6.81 point gain over Self-Consistency CoT.

Toward Conversational Agents with Context and Time Sensitive Long-term Memory integrates a Tabular Chat Database, Classifying Query Type, Chain-of-Tables for Meta-Data Retrieval, and Combining Meta-Data and Semantic Retrieval to handle time-sensitive and ambiguous conversational queries. On the LoCoMo-derived temporal benchmark, Toward Conversational Agents with Context and Time Sensitive Long-term Memory achieves 90.32 average recall vs 31.93 for the best Semantic w MetaD baseline.

TransformerFAM augments Block Sliding Window Attention with Feedback Attention Memory, where Feedback Attention Memory (FAM) tokens attend to and compress block activations while queries jointly attend to BSWA memory segments and past FAM. On PassKey retrieval, TransformerFAM maintains 100% accuracy up to 260k filler tokens, while TransformerBSWA with 12 memory segments collapses after 20k tokens.

Understanding Factual Recall in Transformers via Associative Memories analyzes linear associative memories, MLP associative memories, and a one-layer transformer with multi-head self-attention plus an MLP on a synthetic factual recall task. Understanding Factual Recall in Transformers via Associative Memories proves that storing N random associations requires Θ(N log M) bits and that a single-layer transformer achieves 100% accuracy whenever either self-attention or MLP parameters scale linearly with the number of facts.

CareCall combines a memory management layer, LLM summarizer, and memory-augmented input over HyperCLOVA to store and reuse summaries of users’ Health, Meals, Sleep, Visited Places, and Pets across weekly calls. In deployment to 147 socially isolated adults, CareCall with long-term memory yielded higher Health-detail and Clinical-detail disclosure counts per call than CareCall without memory, and longer average call durations (87.89s vs 75.48s).

LONGMEM augments a frozen GPT-2*-style backbone with a Residual SideNet, Cached Memory Bank, Memory Retrieval and Fusion, and Cross-Network Residual Connections to read and use long-term key–value memories. On ChapterBreak AO3, LONGMEM reaches 40.5% suffix identification accuracy with infinite in-memory context, compared to 28.3% for Memorizing Transformer under the same 1k in-context window.

Empowering Working Memory for Large Language Model Agents introduces a Working Memory Hub, Episodic Buffer, Interaction History Window, Central Processor, and External Environment Interface to give LLM agents persistent, structured working and episodic memory. Empowering Working Memory for Large Language Model Agents is a conceptual blueprint rather than a benchmarked system, so no quantitative MAIN_RESULT against specific baselines is reported.

HimNet combines a GNN Encoder, Node Memory Module, Graph Memory Module, and Graph Decoder to reconstruct graphs via stored normal patterns and score anomalies by reconstruction and approximation errors. On the DD biochemical dataset, HimNet achieves 80.6% AUC compared to 70.6% for PK-iF, a +10.0 point gain over this two-step baseline.

Hierarchical Neural Memory Network (HMNet) stacks multi-level latent memories z1–z3 with Event-write, Up-write, Down-write, Update, and Readout operations driven by Event Sparse Cross Attention. On DSEC-Semantic, HMNet-L3 reaches 57.4 mIoU with event–RGB fusion, improving over a ResNet-50 baseline at 54.1 mIoU while also reducing latency, and on GEN1 HMNet-B1 matches AED at similar mAP with 57% lower latency.

In search of dispersed memories connects associative memory networks, modern Hopfield networks, generative diffusion models, and the denoising loss to show that diffusion score networks encode Hopfield-like energy landscapes in their weights. In search of dispersed memories demonstrates that exact diffusion dynamics reach Pearson correlations of 0.995–0.996 with modern Hopfield iterations on denoising and completion tasks, while classical Hopfield networks reach only 0.700–0.741.

MaLP combines a Dual-Process enhanced Memory (DPeM), Working Memory, Short-Term Memory (STM), Long-Term Memory (LTM), a Coordinator C, and Retriever R to coordinate short- and long-term personalization around a PEFT-tuned LLM. On MaLP’s medical dialogue benchmark with LLaMA-7B, MaLP achieves 69.95% preference classification accuracy and a 91.53% win rate in response generation, improving ROUGE-L profile QA from 29.66 to 33.91 over the LoRA baseline.

MBPTrack combines a Decoupling Feature Propagation Module, BPLocNet, box-prior sampling, and point-to-reference aggregation to track 3D objects from point clouds using temporal memory and size-aware localization. On KITTI, MBPTrack achieves 70.3% Success and 87.9% Precision, improving over CXTrack’s 67.5%/85.3% by +2.8/+2.6.

MemoryBank combines Memory Storage, Memory Retrieval, and a Memory Updating Mechanism to maintain daily conversations, event summaries, and user portraits for long-term personalization. On a 10-day, 15-user simulated dialog benchmark with 194 probing questions, MemoryBank-powered SiliconFriend ChatGPT achieves 0.716 correctness and 0.912 contextual coherence, surpassing SiliconFriend ChatGLM and SiliconFriend BELLE.

Recurrent Neural Networks and Long Short-Term Memory Networks explains how Backpropagation Through Time, LSTM gates and cells, Gated Recurrent Units, bidirectional RNN, and ELMo fit into one dynamical-systems view. Recurrent Neural Networks and Long Short-Term Memory Networks mainly contributes a structured tutorial and survey rather than new benchmark numbers.

Think-in-Memory (TiM) augments an LLM agent with a hash-based Memory Cache, a Hash-based Mapping F(·), and Insert, Forget, Merge organization operations over inductive thoughts. On the Chinese part of the GVD dataset, TiM with ChatGLM raises contextual coherence from 0.428 to 0.665 compared to SiliconFriend.

PMA-Net augments Transformer self-attention with memory banks, prototype keys, and prototype values built via K-Means and k-NN over past activations. On COCO Karpathy test, PMA-Net reaches 131.5 CIDEr under cross-entropy training, a +3.7 gain over a re-trained Transformer baseline.

Working Memory Capacity of ChatGPT probes ChatGPT with verbal and spatial n-back blocks, adding noise, feedback, and chain-of-thought variants to stress working memory. Working Memory Capacity of ChatGPT finds d′ falls to around 1 at n=3 in most conditions and that GPT-4’s verbal n-back capacity far exceeds Bloomz, ChatGLM, and Vicuna baselines.

BayesPCN combines predictive coding, conjugate Bayesian updates over W 0:L, sequential importance sampling, and a diffusion-based forget mechanism to build a hierarchical associative memory that supports continual one-shot writes. On CIFAR10 and Tiny ImageNet hetero-associative tasks, BayesPCN matches offline GPCN with MSE as low as 0.0000 while online GPCN rises to 0.0791 MSE on CIFAR10 mask at sequence length 1024.

Episodic Memory Question Answering combines allocentric top-down semantic features, spatiotemporal memory, and a LingUNet-based question-answering model to localize answers on scene floorplans from egocentric RGB-D tours. On the EMQA benchmark built from Matterport3D tours, Episodic Memory Question Answering with temporal features achieves 29.11 IoU and 62.27 recall on top-down maps, improving over SMNetDecoder by 2.19 IoU and 18.41 recall.

Memory Maze combines an online reinforcement learning environment, an offline dataset, and an offline probing protocol to stress-test long-term memory using Dreamer, Dreamer (TBTT), IMPALA, and supervised probe networks. On Memory 9x9, Dreamer (TBTT) achieves a return of 33.2 compared to 23.4 for IMPALA, while humans reach 26.4 and the oracle reaches 34.8.

Incremental Prompting uses Span-based Event Detection, Episodic Memory Prompting, Experience Replay, and Knowledge Distillation to attach and grow type-specific prompt vectors across tasks. On MAVEN Task 5, Incremental Prompting achieves 50.12% micro F1 versus 39.34% for KT, a +10.78 point gain under class-incremental evaluation.

Knowledge Aware Finetuning (KAFT) augments QA data with relevant context, irrelevant context, counterfactual context, and empty context built from SQuAD 2.0, T-REx, QASC, and TriviaQA, and uses pretrained model’s answer to label irrelevant slices. KAFT achieves up to 24× higher controllability than Noisy Finetuning and up to 6× higher robustness than Noisy Finetuning on PaLM 540B while matching TriviaQA validation accuracy.

Universal Hopfield Networks decompose single-shot associative memory into similarity, separation, and projection, instantiating Hopfield networks, sparse distributed memories, dense associative memories, and modern continuous Hopfield networks within one energy-based framework. Universal Hopfield Networks then compare dot product, Euclidean, Manhattan, and other similarity functions, finding that Manhattan and Euclidean similarity often yield higher retrieval capacity and robustness than dot-product-based modern continuous Hopfield networks on MNIST, CIFAR10, and Tiny ImageNet.

Generalizable Episodic Memory (GEM) combines a parametric network Mθ, implicit memory-based planning, twin back-propagation process, and conservative estimation on single step to generalize episodic returns across trajectories. GEM achieves higher average returns than TD3, SAC, DDPG, and TD3+SIL on MuJoCo tasks like Ant-v2, HalfCheetah-v2, and Humanoid-v2 within 1M environment steps.

Hierarchical Associative Memory organizes fully recurrent Modern Hopfield Networks into layered architectures using Lagrangian functions, hierarchical time scales, and symmetric feedforward–feedback weights as core components. Hierarchical Associative Memory theoretically extends Dense Associative Memories to arbitrary depth and local connectivity, deriving explicit dynamical and energy functions rather than reporting benchmark numbers.

Value-based Episodic Memory (VEM) uses Expectile V-Learning (EVL), Implicit Memory-based Planning, and Generalized Advantage-weighted Learning to learn conservative V-values and plan along offline trajectories. On the D4RL benchmark, VEM attains 87.5 on antmaze-umaze and 128.3 on adroit-hammer-expert, improving over BAIL and CQL on most AntMaze and Adroit tasks.

Episodic Memory Actor Critic (EMAC) augments DDPG with a Memory Module, Episodic-based Experience Replay Prioritization, and a modified critic objective that blends Bellman targets with retrieved Monte Carlo returns. On OpenAI Gym continuous control, EMAC reaches 2236.88 ± 808 average return on Walker2d-v3, beating TD3 (1008.32) by +1228.56 and SAC (1787.28) by +449.6 in the 200000-step small-data regime.

Working Memory Connections for LSTM (LSTM-WM) augments LSTM gates, Working Memory Connections, peephole connections, and the memory cell so gates see a protected projection of the cell state. On PTB character-level language modeling, LSTM-WM achieves 1.299 BPC with TPTB=150 compared to 1.334 BPC for vanilla LSTM.

Emergent Symbol Binding Network (ESBN) combines an LSTM controller, a shared image encoder fe, temporal context normalization, and a two column key value memory to bind abstract variables to concrete image embeddings. ESBN achieves ≥95% test accuracy on same different, RMTS, distribution of three, and identity rules tasks while generalizing to withheld Unicode characters, unlike LSTM, NTM, MNM, Relation Net, Transformer, and PrediNet baselines.

Large Associative Memory Problem rewrites associative memory using coupled feature neurons, memory neurons, an energy function, and Lagrangian functions so that all interactions are pairwise yet recover Dense and modern Hopfield behavior. Large Associative Memory Problem shows that with appropriate choices of activation and Lagrangian functions, the effective dynamics match Dense Associative Memories and modern Hopfield networks, enabling storage of N_mem ∼ min(N_f^{n−1}, N_h) or even exponential in N_f memories without many‑body synapses.

Variational Semantic Memory combines variational prototype inference, variational semantic memory, latent memory m, and an attention-based memory update to build probabilistic class prototypes from long-term semantic knowledge. On miniImageNet 5-way 1-shot with a deep backbone, Variational Semantic Memory reaches 65.72% accuracy versus 64.82% for Tian et al. 2020.

MEMO combines common embeddings, multi head keys and values, recurrent attention, and a halting policy to flexibly chain episodic memories over multiple hops. On joint bAbI 10k, MEMO achieves 0.21% error versus 4.2% for Memory Networks, while also solving long-distance Paired Associative Inference and shortest path tasks.

Self-Attentive Associative Memory (STM) combines Outer Product Attention (OPA), Self-attentive Associative Memory (SAM), Mi-Write, Mr-Read, and Mr-Transfer into a dual item–relational memory system. On the bAbI question answering benchmark, STM attains 0.39 ± 0.18 mean error versus 0.55 ± 0.74 for MNM-p, establishing a new state-of-the-art.

Adaptive Posterior Learning combines an Encoder, Memory store, Memory controller, and Decoder (relational self-attention, relational working memory, or LSTM) to approximate posteriors from a sparse external memory. On Omniglot, Adaptive Posterior Learning achieves 99.9% 5-way 5-shot accuracy, matching MAML and SNAIL, while using fewer than 2 stored examples per class.

Episodic Memory in Lifelong Language Learning combines an example encoder, task decoder, and episodic memory for sparse experience replay and Memory-based Parameter Adaptation (MBPA++). Episodic Memory in Lifelong Language Learning attains 70.6 averaged classification accuracy vs 66.9 for A-GEM and 62.4 QA F1 vs 57.9 for REPLAY on concatenated multi-dataset streams.

Episodic Memory Reader combines a Data Encoder, Memory Encoder (EMR-Independent, EMR-biGRU, EMR-Transformer), Value Network, external memory, and QA solver to learn which streaming items to keep. On TriviaQA, Episodic Memory Reader with EMR-biGRU achieves 57.57 F1 versus 50.10 for LIFO, and on TVQA Episodic Memory Reader reaches about 65% accuracy with 60 memory entries versus roughly 55% for LIFO.

Memory Recall Agent (MRA) integrates a pixel‑input convolutional residual network, an LSTM working memory, a slot‑based episodic memory, an auxiliary contrastive loss, and jumpy backpropagation into a single reinforcement learning agent. Across the Memory Tasks Suite with train/holdout splits, Memory Recall Agent (MRA) attains the highest average human‑normalized performance compared to LSTM‑only IMPALA and other ablations, especially on tasks requiring episodic recall.

On Tiny Episodic Memories in Continual Learning combines Experience Replay, Reservoir Sampling, Ring Buffer, k-Means, and Mean of Features memory writing to jointly train on current-task data and a tiny episodic memory. On Split CIFAR with only 1 example per class, On Tiny Episodic Memories in Continual Learning reaches about 0.56 average accuracy, a +15.6 percentage point gain over FINETUNE and +15 percentage points over EWC.

Working Memory Graphs combines a Core vector, Factor vectors, and persistent Memo vectors processed by a multi layer Transformer to implement shortcut recurrence over past observations. On the Pathfinding task, Working Memory Graphs with Memos nearly matches a full history non recurrent Working Memory Graphs baseline and exceeds a GRU agent by roughly 9.5 percentage points in zero shot quiz accuracy on 24 step episodes.

COG combines Visual processing, Semantic processing, Visual short-term memory, and a Controller so SYS_NAME can parse instructions, attend over images, and maintain working memory. SYS_NAME reaches 96.8% overall accuracy on CLEVR versus 95.5% for CNN+LSTM+RN, while achieving 93.7% on canonical COG and strong zero-shot task generalization.

Working Memory model combines a topic memory, history memory, local memory, genre embedding, and Topic Trace mechanism inside a GRU encoder–decoder to dynamically read and write salient poem context. On Chinese quatrains, Working Memory model achieves BLEU 1.315 and perplexity 86 versus iPoet’s BLEU 0.425 and perplexity 138.

Deep Episodic Memory uses an encoder network E, reconstruction-decoder Dr, prediction-decoder Dp, latent vector V, and a matching and retrieval mechanism to turn raw video into episodic encodings that can be reconstructed and predicted. On ActivityNet, Deep Episodic Memory with PCA achieves 45.55% first-match precision versus 32.31% for ResNet-50 Fisher Vectors, a +13.24 percentage point gain.

Episodic Memory Deep Q-Networks (EMDQN) augments Qθ(s, a) with an inference target S, an episodic memory target H, and a memory table built via random projection and kd-tree lookup. On 57 Atari games at 40M frames, EMDQN achieves a 528.4% mean human-normalized score versus 151.2% for DQN and 144.8% for NEC.

Progressive Memory Banks for Incremental Domain Adaptation augments a BiLSTM with a directly parameterized memory bank, key value memory, and an attention based memory mechanism that is progressively expanded during incremental domain adaptation. On MultiNLI Fic→Gov, Progressive Memory Banks for Incremental Domain Adaptation with memory and vocabulary expansion reaches 67.55% on Fic and 70.82% on Gov, compared to 65.62% and 69.90% for fine tuning with no memory expansion.

Emotional Chatting Machine (ECM) augments a GRU encoder decoder with Emotion Category Embedding, Internal Memory, and External Memory to control emotional content in generated replies. On the Emotional STC dataset, ECM reaches 0.773 emotion accuracy vs 0.724 for Emb and 0.179 for Seq2Seq, while keeping perplexity comparable.

Gradient Episodic Memory stores task examples in episodic memory Mt, constrains updates via inequality constraints on past-task losses, and solves a small quadratic program (GEM QP) to project gradients. Gradient Episodic Memory achieves 0.654 ACC on Incremental CIFAR100 with 5,120 memory slots, compared to 0.508 ACC for iCARL.

Neural SLAM combines an LSTM, Localization and Motion Prediction, Data Association, Measurement Update, and Mapping over an external memory map to guide exploration policies. On 16×16 grid worlds, Neural SLAM achieves 13.732 average reward and 46/50 success episodes, a +6.536 reward gain over A3C-Nav2.

On the Long-Term Memory of Deep Recurrent Networks analyzes Recurrent Arithmetic Circuits, Start-End separation rank, grid tensors, and Tensor Network constructions to quantify how depth affects temporal expressivity. The main result proves depth-2 RACs achieve Start-End separation rank on the order of the multiset coefficient (min{M,R} + T/2 − 1 choose T/2), while depth-1 RACs are limited to rank min{R, M^{T/2}}.

RNN-EM augments a simple recurrent neural network with an external memory Mt, forget gate ft, update gate ut, and a cosine-similarity weight wt to read and write contextual representations. On the ATIS language understanding benchmark, RNN-EM attains 95.25% F1, beating the previous best LSTM result of 94.85% by 0.40 points.