All AI memory papers

FlowState represents agent memory as persistent execution state using a Persistent State Repository, a Historical State Index, and an Active State coordinated by Incremental State Update and Progressive State Access. On MemoryArena, FlowState improves average success rate over Full Context with DeepSeek-V4-Flash by 4.55 percentage points while reducing total token consumption by 43.2%.

ForeDreamer separates factual memory from experiential memory, using a main agent, a memory-processing subagent, an Experience Bank, and a MemGuide–MemTools workspace to process web evidence before forecasting. On Prophet Arena, ForeDreamer achieves an average Brier score of 0.1471 with Qwen3.5-Flash, improving over the Full Text baseline at 0.2059.

PLANFENCE attaches exact parent IDs to plans, validates declared dependencies with authoritative owners, and enforces a dependency-scoped action gate using components like Exact derivation, Declared scope, and Dependency-scoped action gate. In the high-churn AT&T setting, PLANFENCE achieves 0/330 invalid actions with 230.8 ms stall and 8.1 KiB traffic per action, whereas Local replica and Owner-head freshness each issue 330/330 invalid actions.

Memora combines Seed Data Design, Session Simulation, Conversation Generation, and Questions and Evaluation Criteria to build weeks-to-months multi-session benchmarks with evolving memories. Using FAMA on Memora, long-term memory agents like MemoBase fall from 43.60 to 15.18 remembering FAMA between weekly and quarterly settings, revealing severe brittleness under heavy memory mutation.

From Storage to Experience formalizes LLM agent memory into Storage, Reflection, and Experience, detailing mechanisms like Linear, Vector, Structured storage and Introspection, Environment, Coordination reflection. From Storage to Experience’s main result is a unified evolutionary framework that connects fragmented operating system style and cognitive science inspired memory designs into a single roadmap.

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.

GateMem structures shared-memory evaluation around Episodes and Memory State, Checkpoints and Governance Categories, and a multiplicative Memory Governance Score that couples utility, access violations, and forgetting failures. GateMem shows, for example, GPT-5.4 with LONG-CONTEXT reaches 80.1% MGS on the Medical domain while RAG-NAIVE only achieves 44.7%, yet even LONG-CONTEXT still leaks unauthorized or deleted information.

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.

Graph-based Agent Memory organizes agent memory into Knowledge vs Experience Memory, Short-term vs Long-term Memory, and Non-structural vs Structural Memory tied to graph implementations. It unifies memory extraction, storage, retrieval, and evolution into a single lifecycle, mapping over 70 named systems like MemGPT, GraphRAG, and Zep into a coherent design space.

Graph-Based Personalized Memory for LLM Agents organizes memory representation, memory evolution, memory retrieval, and memory evaluation into a single lifecycle for LLM personalization. This unified taxonomy clarifies how graph-based memory supports long-horizon personalized assistants across benchmarks like LoCoMo, PersonaMem-v2, and EngramaBench.

GraphMemix combines candidate graph construction, evidence utility and activation costs, and forest optimization to organize long-term multimodal memories as a query-conditioned evidence forest. On four personal multimodal memory benchmarks with Qwen3-VL-8B-Instruct, GraphMemix reaches 61.55% macro Judge Accuracy on ATM, Mem-Gallery, MemEye, and H2HMem, improving over UniversalRAG’s 49.80% by 11.75 percentage points.

Grounding Agent Memory wires a task agent, curator agent, memory index, trajectory distillation, and environment-probing curation into a propose–probe–commit pipeline for enterprise agents. On CLBench with schema drift, Grounding Agent Memory achieves 73% pass rate and 22.60 total reward versus 39% and 8.60 for GHCP (No Memory).

MemSuit combines Adaptive Entry Decomposition, Utility-Aware Self-Distillation, and Entry-Aware Retrieval Calibration to turn raw conversations into compact, query-useful memory entries and aligned embeddings. On LoCoMo with Llama-3.1-8B-Instruct, MemSuit reaches 35.56 average F1 and 27.73 BLEU, beating SimpleMem’s 30.09 F1 and 22.24 BLEU.

GroupMemBench combines Graph-grounded Message Synthesis, Path-Based Context Sampling, Message Generation, Adversarial Question Synthesis, and a Solve–Judge–Refine Loop to stress-test group memory in LLM agents. On GroupMemBench, Hindsight reaches 46.01% average accuracy while a simple BM25 baseline attains 43.22%, revealing that current memory ingestion pipelines fail to preserve crucial multi-user structure.

H2HMem constructs dyadic and multi-party multimodal conversations via a Participant Profile Generation, Scenario Construction, Image Collection and Human Refinement, Image Captioning and Dialogue Generation, and Question-Answer Pairs Construction pipeline, then evaluates agents on recall, reasoning, and application tasks. On the H2HMem benchmark, the best configuration with A-Mem on GPT-4.1-Nano reaches an overall LLM-as-Judge score of 0.5757, revealing large gaps versus ideal memory behavior.

Harness the Memory instruments a Unified Evaluation Harness, External Memory Substrate Families, Internal Memory Substrate Families, Retrieval Depth Sweep, and Scalability Study to compare 11 memory substrates under identical agents. Harness the Memory shows, for example, that M7 Distilled Strategies reaches 32.1% TSR on ALFWorld-unseen with QWEN3-32B-AWQ, a +9.7 percentage point gain over the NoMem baseline at only 1.23× latency.

HERO organizes dialogue into Episodic Traces, Episodic Units, Episodic Cues, Profile Insights, and Profile Cues inside a heterogeneous memory graph, then runs profile-aware cue activation and Personalized PageRank for retrieval. On LoCoMo, HERO attains 56.06% F1 and 87.99% accuracy, beating EverMemOS by +10.58 F1 and +3.96 accuracy.

HiGMem organizes dialogue history into a Turn Layer, Event Layer, Memory Construction Mechanism, and LLM-Guided Retrieval that use event summaries as semantic anchors for turn selection. On LoCoMo10, HiGMem reaches 0.78 adversarial F1 and 0.49 single-hop F1, beating A-Mem while using an order of magnitude fewer retrieved turns.

HiMem builds Episode Memory, Note Memory, Knowledge Alignment, and Memory Reconsolidation to construct, retrieve, and update long-term dialogue memories hierarchically. On the LoCoMo benchmark, HiMem achieves an overall 80.71% GPT-Score, beating the best baseline SeCom at 69.03% by 11.68 points.

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.

HippoCamp evaluates contextual agents on realistic personal file systems using structured trajectories with search, perception, and reasoning capability tags over 42.4 GB of multimodal data. On HippoCamp, ChatGPT Agent Mode reaches 48.3% profiling accuracy and 62.8% factual retention accuracy, while RAG and search agents lag far behind.

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.

HYPERSKILL organizes experiential memory as a hypergraph of Nodes, Hyperedges, Dual-path retrieval, Co-occurrence ranking, and HYPERSKILL Memory Evolving to store and reuse subtask and skill knowledge. On GAIA with GPT-4o, HYPERSKILL achieves 44.24% success rate, improving over PlugMem’s 41.94% by +2.30 points.