All AI memory papers

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.

LazyMem stores raw histories and at query time runs Hybrid Retrieval, History Windowing, and a trained Memory-Processing Model to build compact, query-conditioned memory for an answer LLM. On LongMemEval, LazyMem-4B achieves 0.85 LLM-judge accuracy with 213 answer-context memory tokens, a 0.03 gain over StructMem while using 21.0× fewer tokens.

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.

LoCoMo-Plus evaluates cognitive memory using an implicit Cue–Trigger Query Construction pipeline plus Semantic Filtering, Cue Memory Elicitation Validation, and insertion into LoCoMo dialogues. On LoCoMo-Plus, even strong systems like gemini-2.5-pro reach only 26.06% cognitive accuracy versus 71.78% factual accuracy on LoCoMo, exposing a large unresolved gap.

LongMemEval-V2 evaluates long-term agent memory via context gathering over massive web-agent haystacks, using components like AgentRunbook-R, AgentRunbook-C, raw state slice pools, state transition event pools, and procedure and hint note pools. On LME-V2-Medium, AgentRunbook-C achieves 70.1% overall accuracy, a +24.2 point gain over the strongest RAG baseline with trajectory notes (45.9%).

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.

LycheeMemory V2 combines Online Semantic Segmentation, Segment-Level Memory Encoding, Structured Evidence Organization, and Plan-Guided Multi-Route Retrieval to batch coherent dialogue segments into typed, indexed records. On LoCoMo, LycheeMemory V2 achieves 89.22% overall accuracy versus 68.83% for A-Mem, and on LongMemEval-S it reaches 92.20% versus 71.60% for A-Mem, with up to 7.2× fewer construction tokens.

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.

Mandol combines a hierarchical memory model, SemanticMap, SemanticGraph, and a quantitative query mechanism to unify agent memory and retrieval. On LoCoMo with GPT-4.1-mini, Mandol achieves 92.21% overall accuracy, 0.24 percentage points higher than EverMemOS while using fewer tokens.

MAPLE orchestrates a central Agent with dedicated Memory, Learning, and Personalization sub-agents that run on different timescales but share explicit user models. On the MAPLE-Personas benchmark, MAPLE achieves a mean judge personalization score of 4.78 vs 4.17 for the Baseline and raises trait incorporation from 45% to 75%.

MELD merges distributed wiki brains using a merge decision procedure, Patch, status CRDT, and semantic publish subscribe over a typed-link memory substrate. On HotpotQA distractor, MELD’s decentralized merge achieves recall@5 = 0.630 vs 0.619 for a centralized store and 0.595 for naive union, with about 11% less live storage than union.

Mem2ActBench integrates Heterogeneous Data Integration, Fact Extraction and Grouping, Memory Evolution Chain Construction, and Memory-anchored Q&A Construction to create long, interruption-heavy sessions with grounded tool calls. On Mem2ActBench, oracle retrieval reaches F1 53.8 while the best passive hybrid retriever at k=5 reaches only 30.7, revealing a 23.1 F1 gap in memory utilization.

MemAdapter combines a Generative Subgraph Retriever, Anchored Alignment Module, Target Alignment Module, Unified Memory Space, and Agent Model to turn diverse memory states into explicit evidence subgraphs. On NarrativeQA with Qwen2.5-7B, MemAdapter achieves F1 61.59 compared to 56.91 for MemoryLLM and 50.01 for Mem0, while completing cross-paradigm alignment in about 13 minutes with less than 5% of training compute.

MemAudit combines Counterfactual Memory Influence Score, Memory Consistency Graph, and a fused Detoxification Score to rank and remove suspicious memories in MINJA-poisoned agents. On MINJA QA with GPT-4o, MemAudit reduces attack success rate from 70.0% to 0.0%, beating random deletion, retrieval-frequency deletion, and nearest-neighbor contradiction filtering.

MemBuilder trains a Qwen3-4B-based agent to coordinate Core Memory, Episodic Memory, Semantic Memory, and Procedural Memory using dense session-level rewards and contribution-aware gradient weighting. On LoCoMo, MemBuilder achieves 84.23% accuracy, beating MIRIX at 77.48% and Claude 4.5 Sonnet memory construction at 82.61%.

MemConflict builds structured user profiles, simulates long-horizon timelines, injects conflicts, and evaluates systems via black-box Answer Accuracy and white-box Support Evidence Hit and Support Rank Score. On the MemConflict benchmark, MemOS attains 0.5539 average Answer Accuracy, 0.6710 SEH@3, and 0.5879 SRS, revealing large gaps versus systems like LangMem at 0.2822 AA.

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.

MemForest combines parallel extraction, a shared memory substrate of canonical facts, scoped MemTrees, and forest recall plus tree browse to maintain temporal agent memory efficiently. On LongMemEval-S, MemForest with Qwen3-30B-A3B-Instruct-2507 achieves 81.8% pass@1 overall, 14.8 points above EverMemOS, while its input-normalized build rate is 6.0× higher.

Mem-Gallery organizes multi-session, multimodal conversations and tasks via Benchmark Construction, a unified Conversational Environment for Memory, and an Evaluation Framework spanning extraction, reasoning, and knowledge management. On Mem-Gallery, MuRAG reaches 0.6966 overall F1 and 0.8856 LLM-Judge on visual-centric search, surpassing NaiveRAG by 0.0992 overall F1.

MemGround evaluates long-term memory via Surface State Memory, Temporal Associative Memory, and Reasoning-Based Memory in TRPG, No Case Should Remain Unsolved, and Type Help scenarios. MemGround’s main result shows frontier GPT-5.2 reaches 51.51% QA Overall in TRPG but only 23.61% QA Overall in Type Help, revealing severe reasoning degradation versus closed-source baselines like Gemini-3-Pro-Preview.

MEMGUARD decomposes conversations into type-specific memory atoms via Type-Aware Knowledge Decomposition, Self-Verified Extraction, Relational Knowledge Graph, and Query-Adaptive Type Routing stored in type-isolated memory. On HaluMem, MEMGUARD attains 89.53% anti-hallucination accuracy and 94.15 F1, improving anti-hallucination by 28.27 percentage points over MemOS.