All AI memory papers

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.

DreamBench-SWE wraps a fixed wake agent with a raw episode log, derived memory store, maintenance pipeline, and retrieval gate to test multi-session memory hygiene in software repositories. On 60 traps and 180 S3 cells, the strongest verbatim baseline B5 reaches 89/180 Pass@1 while the hybrid reference probe reaches 95/180, but the clustered P1 comparison fails to reject (p=0.518).

Dual-Layer Agentic Memory combines a fast write router, small-to-large cost-aware routing cascade, operational memory taxonomy, and write-back consolidation to manage parametric and external memory. On the streaming ZsRE benchmark, Dual-Layer Agentic Memory’s Write Router retains 86.35% QA EM versus 87.79% for Full Store, while cutting storage to 77.00% and reducing 8B routing compute by up to 39.7%.

DynamicMem builds 15‑month, multi‑app trajectories using Multi-timescale user profile construction, Intent-conditioned event chain generation, and State-consistent multi-app log generation to stress-test long-horizon memory. DynamicMem’s main result shows State Completion declines by up to 26.5 points for A-Mem from checkpoint C1 to C5, while Personalized Service scores remain stable or improve, revealing hidden failure modes in memory systems.

ECHO orchestrates Immutable events, Typed projection, Bitemporal ledger, Candidate discovery, Provenance closure, and Derive then realize into a governed memory plane for long-horizon agents. On LongMemEval-S, ECHO reaches 97.60% Hit@10 and 88.84% turn Recall@5, while a matched QA sample shows Mem0 OSS at 64.84% versus ECHO’s 41.76% (−23.08 percentage points).

EchoPath wraps GUI agents with ActionLens, an EchoPath Memory Repository, a Replay Module, Memory Consolidation, and an Image-based Target-Reaiming (IBTR) algorithm to turn validated trajectories into callable memories. On OSWorld-Verified tasks, EchoPath achieves 91.2% second-pass success with codex-gpt-5.5-medium while reducing median token usage from about 572k to 20,370 compared to Synapse’s 586,386.

EdgeMem stores verbatim interaction turns and organizes them via a Multi-anchor Hypergraph Construction, Time Sub-Hypergraph, Co-occurrence Sub-Hypergraph, and Episode Sub-Hypergraph with deterministic retrieval. On LoCoMo, EdgeMem reaches 61.01 strict-judge accuracy versus 58.70 for CompassMem while requiring 0 construction tokens and only 1,000 tokens per question.

EM2Mem organizes long videos into Event-Centric Multimodal Memory Cells, Temporal Context Views, and event-linked Episodic and Semantic Graphs for align-then-retrieve reasoning. On Video-MME (L), EM2Mem reaches 76.8% average accuracy, beating the strongest memory baseline WorldMM† at 73.1% and cutting total inference tokens by 63.66%.

EMemBench builds on Interaction episode, Benchmark generator, Structured logging, State reconstruction, and Programmatic QA generation to turn each agent’s own game trajectory into verifiable episodic-memory questions. On text-only games, EMemBench shows A-MEM with Qwen3-32B achieves 51.9% Overall ACC compared to 44.9% for the Qwen3-32B in-context baseline (+7.0 points).

E-mem combines a Master Agent, multiple Assistant Agents, a Multi-Pathway Routing Mechanism, Sliding Window Segmentation with Overlap, and Episodic Memory Context Retention and Isolation to reconstruct native contexts instead of compressing them. On the LoCoMo benchmark, E-mem achieves 54.17 F1 with GPT-4o-mini, beating GAM by 8.86 F1 while reducing normalized token cost from 169100 to 3621.

EngramaBench evaluates Engrama using graph Memory activation, Neighborhood retrieval, Structured summary, and a Typed answer layer over multi-session persona histories. On EngramaBench full v1, Engrama reaches 0.6532 cross space versus 0.6291 for GPT-4o full-context while using only $0.67 vs $3.33 in query-time cost.

ENPMR-Bench evaluates Emotional Need-aware Proactive Memory Retrieval by tying User Profile Generation, Structured Memory Retrieval Guidelines, Emotional Support Dialogue Generation, and Memory Retrieval Task into a controlled benchmark. On ENPMR-Bench, Qwen3-Embedding-8B reaches only 46.42% Recall@10, while DeepSeek-V3 empathy rises from 4.38 to 4.91 when ENPMR-Bench supplies golden memories instead of retrieved ones.

EnSIMem organizes long dialogues via Theme-coherent episodic memory construction, Dialogue-grounded entity-property indexing, Granularity-controlled property alignment, and Requirement-aware evidence localization into an entity structured index linked to original episodes. On LongMemEval, EnSIMem reaches 92.8% average accuracy versus 87.4% for MEMORA (P), and on LoCoMo EnSIMem reaches 90.6% versus 86.3%.

EP-Mem combines a User Pre-configured Privacy Policy File, Elastic Privacy Memory Unit, EP-Mem Sidecar Overlay, Evolvable EP-Mem Privacy Engine, and Memory Material Preprocessing to enforce social relationship-aware disclosure across long-term agent memory. On EP-Bench Task 3, EP-Mem+Evo achieves PB=0.136 versus 0.738 for NoEngine_evid while improving MIQ from 3.30 to 4.32.

ER-MIA combines Content-Based MIAs, Question-Targeted MIAs, and an attack Arsenal of primitive and ensemble strategies to poison similarity-based retrieval in long-term memory systems. On the LoCoMo benchmark, ER-MIA’s ensembles cut Mem0’s overall F1 from 23.60% to 2.87% (-87.6%) and reduce single attacks like Harsh Instruction by up to -71.5% F1.

ERRAND combines drift-aware beliefs, a single-peaked errand index, a deadband gate, and a versioned lifecycle to decide when an agent should spend actions revalidating memory instead of working. On the main scripted tool-use world, ERRAND reaches 71.9% ITT success at b=12 compared to 61.9% for eager revalidation (+10.0pp) while spending only 11.0% of steps when uncapped.

ES-MemEval evaluates conversational agents’ long-term memory via Question Answering, Summarization, and Dialogue Generation on the EvoEmo dataset, which combines User Profile Construction, Event Timeline Expansion, and Chat Data Generation. On ES-MemEval QA, Mistral-24B + RAG reaches 18.8% overall F1 and 1.27 LLM-as-Judge score, improving over base Mistral-24B and revealing how explicit memory and retrieval reshape performance.

EvalMem decomposes long-term memory systems into Encoding Examiner, Retrieval Examiner, Generation Examiner, and an Attribution Agent that assigns 11 fine-grained defect codes. EvalMem shows retrieval defects reach 22.1% on LOCOMO, and a MemWiki auxiliary index guided by this diagnosis improves mean accuracy by 2.5 percentage points over seven memory systems.

EvoGraph-Mem maintains an editable Insight Graph, Query Graph, Utility-Aware Insight Retrieval, and a Graph Controller for Memory Correction to keep long-term agent memories reliable. On PDDL, HotpotQA, and FEVER, EvoGraph-Mem reaches a 47.47 average score with GPT-4o-mini, +4.24 over G-Memory’s 43.23.

EvolMem combines Topic-Initiated Generation, Narrative-Inspired Transformation, Filtering, and Challenge Injection to build 1,600 multi-session dialogues covering declarative and non-declarative memory. On EvolMem, Gemini-3-Pro achieves 67.47% overall memory performance, ahead of GPT-5.1 at 52.82%, while agents like MemoryOS fail to surpass their DeepSeek-V3.2 base.

EvoMemBench evaluates self-evolving agent memory using four settings and six datasets spanning in-episode and cross-episode, knowledge and execution evolution. EvoMemBench’s main result is that explicit memory only helps reliably when context is insufficient, tasks are hard, or stored experience matches the decision process, while Gemini-3-Flash and other long-context baselines often match or beat 15 memory methods.

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.

FadeMem manages agent memory using a Dual-Layer Memory Architecture, Biologically-Inspired Forgetting Curves, Memory Conflict Resolution, and Adaptive Memory Fusion to selectively retain and consolidate information. On LoCoMo, FadeMem attains 29.43 multi-hop F1 and 0.45 SRR, improving over Mem0’s 28.37 F1 while cutting storage by 45%.

FinPerMA combines Persona Synthesis, Event Timeline Construction, the three-layer Impact Model, Dialogue Synthesis, and Evaluation Task Design to generate frozen, event-grounded investor trajectories for memory benchmarking. On the FinPerMA v8gold corpus, full-context Qwen-3.8 reaches 46.9% overall accuracy and 38.7% MCQ, while retrieval-based memory on Qwen3.7-Max recovers ≈88% of the no-memory–full-context gap with only 1.40k tokens of context.