All AI memory papers

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.

AutoMem composes long-term memory from modular Encode, Store, Retrieve, and Manage components, optimized via Experience-Guided Architecture Search and Failure-Guided Module Diagnosis. On GAIA with Qwen3.5-122B-A10B, AutoMem achieves 71.5% accuracy vs 67.8% for MemoryBank, and reduces token cost by 14.3%.

AutoSkill composes prompt-driven modules Query Rewriting, Hybrid Skill Retrieval, Skill Extraction, Skill Management Decision, and Versioned Skill Merging to externalize user preferences and workflows into SKILL.md artifacts stored in a SkillBank. AutoSkill builds a multilingual SkillBank with 1,858 skills across four WildChat-1M subsets, showing that explicit skill artifacts can be continuously refined (e.g., professional_text_rewrite at version 0.1.34) without any parameter updates.

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.

Bad Memory evaluates how auto-loaded and referenced workspace files like CLAUDE.md, AGENTS.md, and behaviors.md affect agent behavior across sessions using a synthetic workspace, probe sessions, and adversarial goals. Bad Memory finds attack success rates as high as 100% and persistence rates up to 100% across Claude Code and Codex agents, revealing distinct failure modes in memory-based prompt injection.

BeliefMem maintains an external belief-based memory bank with Add, Merge, and Belief-aware Retrieval over attribute-level hypotheses. On LoCoMo, BeliefMem reaches 42.38 F1 with GPT-4o-mini, beating Mem0’s 40.99 F1, and on ALFWorld it attains 59.88% success rate vs ReadAgent’s 54.03% (+5.85).

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.

Threader organizes raw dialogue using Incremental Topic-Coherent Segmentation, Multi-View Segment Representation, and Multi-Signal Evidence Retrieval to access long-horizon memory without LLM rewriting. On LongMemEval with GPT-5-mini, Threader achieves 90.60% overall accuracy and 93.59% Knowledge-Update accuracy, beating EverMemOS by 6.84 and 5.80 points respectively.

CAMA combines Neuro-Symbolic Evidence Assignment, Effective Independent Evidence Estimation, Factor-Level Conflict Arbitration, and Active Independent-Evidence Recovery to decouple correlated memories into latent evidence slots and recover missing sources. On LongMemEval with DeepSeek V4 Flash, CAMA reaches 87.9 EM compared to MADAM RAG’s 86.8 EM while also cutting Replication Sensitivity from 15.3 to 7.8.

xMemory organises agent memory via segments, memory components, groups, and dynamic maintenance with a sparsity plus semantic faithfulness objective. On LoCoMo with GPT-5 nano, xMemory reaches 38.71 BLEU and 50.00 F1 versus Nemori’s 36.65 BLEU and 48.17 F1, while cutting tokens per query from 9154.76 to 6581.20.

RHELM builds 10 persona trajectories with Profile Generation, LOOP, External Sources, Dialogue Synthesis, and Question Curation to create 11,764 turns and 2,180 documents over 629 days. On the RHELM benchmark, Claude Opus 4.5 with RAG at k = 20 reaches 38.1 average score with external data, revealing substantial gaps in long term memory compared to simpler daily life benchmarks.

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.

CABLE augments long-term memory systems with Antecedent-Oriented Query Generation, Dual Retrieval, Overlap Subtraction, and Verification and Graph Update to build a sparse directed graph of complementary links. On MA-LongMemEval with A-MEM and Qwen3.5-27B, CABLE raises the mean LLM-judge score from 59.33% to 65.33%, a +6.00 percentage point gain over the A-MEM baseline.

StateMem represents conversations as structured StateStore entries built by a TurnEncoder, updated via deterministic Rechecker passes, and queried through a guided test-time recomputation prompt. On StateMemBench, StateMem reaches 0.363 gold rate on DeepSeek-V4-Flash, a +0.158 gain over the best retrieval baseline (Dense at 0.205) and +0.164 over long-context (0.149).

CapMem evaluates caption-based episodic memory using CleanQA, Direct VideoQA, CaptionQA, and a retrieve-and-verify harness over 75 egocentric videos and 1,000 QA pairs. On CapMem, 30-second CaptionQA lifts Qwen3.5-27B from 45.0% VideoQA accuracy to 49.3%, and GPT-5.2 from 37.3% to 52.0%, with long-video gains up to +15.9 points.

CAPTURE combines a preference-hypothesis extractor, authenticity gate, multi-timescale ledger, quarantine, and safety-bounded selector with causal audit to track latent user state over long horizons. On the D-PREFGUARD benchmark, CAPTURE achieves 71.5% win rate and 11.5% poisoning versus 69.3% and 15.9% for Sup. Transformer-∆t.

CAST builds views, scenes, character profiles, a semantic index, and an episodic index to organize dialogue into person-conditioned event structures. CAST reaches 62.02 F1 and 81.21 J on LOCOMO open questions, beating Zep by +13.24 F1 and +8.32 J and vanilla RAG by +25.15 F1 and +24.61 J.

MERIT combines a Deterministic Error Classifier, Online Dual-Polarity Memory, Error-Typed Hybrid Retriever, and Type-Reliability-Aware Retrieval Variant to reuse oracle-verified corrections across queries. On Spider, MERIT reaches 69.79% execution accuracy versus 66.34% for Iterative repair using the same Qwen2.5-7B-Instruct backbone and initial predictions.

Chain-of-Memory (CoM) combines Memory Construction and Retrieval, Dynamic Memory Chain Evolution, State-Aware Gating Evolution, and Adaptive Path Truncation to turn flat retrieved turns into coherent reasoning chains. On LongMemEval with Qwen3-32B, Chain-of-Memory (CoM) achieves 76.40% total accuracy versus 66.00% for RAG (turn), while using only 8.8k tokens compared to 119.6k for Full-Context.

CHIME separates experience into a Hierarchical Memory Bank with a Plan Bank, Execution Bank, a Credit Attribution Gate, and Credit-Aware Memory Evolution to update memories only where credit belongs. On the DeepSeek-V4-Flash backbone, CHIME attains 39.01% average eval accuracy across τ²-Bench, VitaBench, BrowseComp-ZH, and BFCL-v4, beating A-MapReduce (35.33%) by +3.68 percentage points.

ChronoMem wraps ADK’s LocalMemoryService, SQLite memory store, Version Index, and Reranker module into a semantic version-control layer that snapshots agent memory on every write and rolls back via natural-language queries. ChronoMem reaches 55.1% rollback-consistent QA accuracy on MemoryAgentBench with Qwen2.5-7B, compared to 35.5% for the RAG-only baseline.