All AI memory papers

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.

MEMO builds query-specific working memory by combining a query-conditioned evidence extractor, unit-materialization, a memory manager, and deterministic working memory building into a multimodal text plus visual pipeline. Under a 128-token budget on 2WikiMultiHopQA with Qwen3-VL-32B, MEMO reaches 73.91 F1, a +17.65 gain over Text-only memory at 56.26 F1.

MemOps decomposes long-horizon memory into structured Remember, Forget, Update, Reflect, and TrajectoryOps traces with six probe types like OperationTrace and StateTransition. On the MemOps benchmark, Claude-Sonnet-4.5 attains 0.916 adjacent accuracy versus Temp-LoRA’s 0.162, exposing deep reliability gaps across memory paradigms.

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.

Memory-Augmented Compression builds a memory bank, uses a memory retriever, performs memory-augmented prefill, and runs memory-guided compressed inference to support short Chain-of-Thought reasoning. On GSM8K, Memory-Augmented Compression with CoD achieves 89.3% accuracy versus 67.9% for CoD alone, while keeping 1.49× lower latency than standard CoT.

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.

MemoryCPT combines Query-agnostic Distillation, Query-aware Retrieval and Summarization, RRF-based Retrieval, and a GRPO-trained Fine Summary adapter on top of Qwen2.5-7B or Llama-3.2-3B. On LoCoMo, MemoryCPT with Qwen-2.5-7B achieves F1 0.479 versus BudgetMem’s 0.373, while reducing Cost from 24.11 to 4.31.

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.

Memory Boundary-Aware Router combines Behavioral Scientific Knowledge Boundary Characterization, Knowledge-Circuit View of Internal Knowledge Boundaries, External Boundary-Aware Data Router, and Internal Boundary-Aware Parameter Router to control conditional memory inside Transformer blocks. On BioProBench Error Correction, Memory Boundary-Aware Router reaches 0.63 accuracy with Qwen3-8B, surpassing the 0.62 accuracy of Qwen3-8B-Memory-LoRA.

MemoryLake organizes confirmed conclusions, supporting evidence, and reusable experience into separate tracks with presence policies, using gpt-5-mini and bge-m3 for consolidation, retrieval, and bounded prompt assembly. On the shared MemoryArena sets, MemoryLake reaches a 20.5% equal-weight macro-average SR compared to the 13.6% best comparator, with SR 9/40 in mathematics, 12/20 in physics, and 4/20 in progressive retrieval.

Memory Poisoning Attack and Defense on Memory Based LLM-Agents combines Input Output Moderation, Memory Sanitization with trust-aware retrieval, bridging steps, and indication prompts to study and harden long-term memory in EHR agents. On MIMIC-III with GPT-4o-mini and Llama-3.1-8B-Instruct, Memory Poisoning Attack and Defense on Memory Based LLM-Agents shows that adding realistic initial memory drops GPT-4o-mini ASR from 62% to 6.67% and ISR from 100% to 26.67% compared to the empty-memory baseline.

MemoryRewardBench constructs paired memory management trajectories across Sequential Pattern, Parallelism Pattern, Mixed Pattern, long-context reasoning, multi-turn dialogue, and long-form generation to test 13 reward models. MemoryRewardBench shows Claude-Opus-4.5 at 74.75 average accuracy and GLM4.5-106A12B at 68.21 on its 2,400-example benchmark, exposing generational gains over larger predecessors.

MemPrivacy combines a Local MemPrivacy Model, Four-Level Privacy Taxonomy, Local Secure Mapping DB, and Uplink Desensitization / Downlink Restoration to sanitize edge inputs and restore them after cloud processing. On MemPrivacy-Bench, MemPrivacy-4B-RL achieves 85.97 F1 vs 68.99 for GPT-5.2, with ~2.05s latency instead of 4.62s.

MEMPROBE combines Simulated Users and Leak-Controlled Tasks, Simulation Rollout, and Recovery Scoring to audit what user state memory agents actually preserve. On 1,550 hidden user-state targets, MEMPROBE finds category-balanced reconstruction only around 0.624 for longctx_full under full-store access and as low as 0.473 for mem0 under top-k retrieval, even though all systems reach ~99.9% task completion.

MemReader combines MemReader-4B, MemReader-0.6B, a ReAct-style trajectory policy, and a memory-operation tool space to actively judge value, ambiguity, and completeness before writing memories. On LongMemEval, MemReader-4B-GRPO achieves 83.00% Overall vs 83.00% for EverMemOS, with 922 vs 2800 average tokens per extraction.

MemReranker connects Qwen3-Reranker with multi-teacher label generation, BCE pointwise distillation, InfoNCE contrastive fine-tuning, and multi-turn dialogue data engineering to build reasoning-aware memory rerankers. On LongMemEval, MemReranker-4B attains MAP 0.8043 versus 0.7259 for Gemini-3-Flash, while MemReranker-0.6B matches or exceeds GPT-4o-mini at ∼8× lower latency.

MemRetriever orchestrates parallel_search, serial_search, reflection_and_denoise, and a GRPO-trained search policy to iteratively build a compact evidence pool from long-term memory. On LOCOMO, MemRetriever-4B-RL reaches 0.830 llm-as-judge accuracy, +0.025 over MemRetriever-4B-SFT under the same retrieval pipeline.

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.

MemRouter combines a Memory Router Architecture, Memory Store and Retrieval, and Answer Agent built on a frozen Qwen2.5-7B backbone to make turn-level memory admission decisions in embedding space. On LoCoMo, MemRouter achieves 52.0 overall F1 with Qwen2.5-7B-Instruct, beating Memory-R1-GRPO at 43.1 F1 under a like-for-like backbone and reaching 55.5 F1 when swapping only the answer agent to Qwen3.5-35B-A3B.

MemSecBench links Lifecycle Task Packages, Build-MemSecBench-Case Skill, Lifecycle Evaluation Workflow, Evidence-Based LLM Judging, and a 24-configuration matrix to trace malicious semantics through memory systems. MemSecBench reports 50.3% End-to-End Attack Success Rate and 56.1% Selective Repair Success Rate across 310 cases and 24 configurations, contrasting Native with Mem0, Mem0-Graph, and A-MEM.

MemSentry evaluates persistent memory operations via a 9-step pipeline using components like Source Trust, Semantic Classification, Attack Radius, and Access Risk over a dependency DAG. On a 1,000-scenario GPT-4 dataset, MemSentry with SBERT+LR achieves 91.7% accuracy and 0.908 macro-F1, detecting 100% external quarantine-class threats compared to Regex, TF-IDF+SVM, and SetFit.

MemSyco-Bench evaluates agent memory use across Objective Fact Judgment, Contextual Scope Control, Memory-Evidence Conflict, Personalized Memory Use, and Valid Memory Selection, using a structured memory-decision schema, question instantiation, long-term dialogue simulation, and multi-stage quality validation pipeline. On MemSyco-Bench, existing memory systems like SuperMemory and MemGPT frequently degrade accuracy (e.g., -23.12 points) and increase sycophancy rates (e.g., +37.24 points) compared to no-memory baselines.

MemToC builds a controlled arbitration pipeline from ToolHop pool construction and filtering, Distractor generation, Correctness-conditioned evaluation, and Annotator quality control to separate source correctness from source preference. On MemToC, cross-fitted SFT lifts Llama-3.1-8B-Instruct correct-answer retention from 17.1% to 31.6% while keeping correct-tool following at 92.8–93.1%, and DPO achieves 22.3% retention with 92.1% correct-tool following.