All AI memory papers

Chronos decomposes dialogue into structured events via the Event Extraction pipeline, stores them in dual event calendar and turn calendar indexes, and uses Dynamic Prompting, Initial Retrieval, and the Chronos Agent for temporal-aware tool-calling. On LongMemEvalS, Chronos Low reaches 92.60% overall accuracy and Chronos High 95.60%, beating EmergenceMem Internal by 7.67 percentage points and Mastra’s OM by 3.02 points.

ClawVM manages agent state as typed pages via the SessionPageTable, RepresentationSelector, FaultObserver, WritebackJournal, and ClawVMEngine inside the agent harness. Across four OpenClaw-derived workloads and six token budgets, ClawVM cuts explicit faults from 67.8 (retrieval baseline) and 1.5 (Compaction-Hybrid) to 0.0 while adding median <50 μs policy-engine overhead per turn.

CLONEMEM builds AI clone memory from Persona and Macro-Level Life Arcs, Meso-Level Phase Generation, Micro-Level Digital Trace Generation, and Evaluation Question Construction over synthetic diaries, posts, and emails. On CLONEMEM, flat retrieval achieves Recall-Any-Any 0.6103 at k=20 versus 0.3913 for A-Mem, revealing a validity–fidelity trade-off in existing memory systems.

CMI-Mem manages long-term dialogue memory via a Structured Multi-dimensional Memory Architecture, Action-conditioned CMI Reward, and Reinforcement Training with GRPO that jointly optimize what to store, update, or discard. On LoCoMo, CMI-Mem-4B achieves 72.1% accuracy compared to 68.1% for MemBuilder-RL-4B, and on MemoryAgentBench CMI-Mem-8B reaches 56.7 vs 51.7 for MemBuilder-RL-8B.

Codebase-Memory parses repositories with a multi-pass pipeline using the Parse stage, Build stage, Serve stage, FunctionRegistry, Louvain communities, and MCP tool interface to build a persistent SQLite knowledge graph. On a 31-language benchmark, Codebase-Memory reaches 0.83 quality versus 0.92 for an Explorer Agent while using ten times fewer tokens and 2.1 times fewer tool calls.

CoEvo-Mem routes queries via the Self-Routed Query Rewriter, a Prior-Guided Residual Router, a Q-Weighted Hybrid Retriever, and a Typed Relational Memory Graph with Trajectory-Conditioned Memory Update. CoEvo-Mem attains 68.33% accuracy on GPQA Diamond, beating MemRL and MemQ at 60.83% by +7.50 percentage points.

CoMem combines Private Experience Sedimentation, Collective Wisdom Curation, Parallel Dual-Stream Retrieval, and a strict empirical promotion gate into a two-tier group memory for multi-agent systems. On ALFWorld with MacNet, CoMem reaches 89.55% success versus the 79.85% No-memory baseline, and also lifts PDDL success from 60.78% to 70.19%.

RSM-full combines Cosine-gated max-member merge, Atom-aware grouped context assembly, |v⊤₁,m q| retrieval, and higher-rank storage variants into an online clustered-memory pipeline for LLM agents. On AMA-Bench, RSM-full achieves 0.311 average correctness at ~4k tokens, beating Online K-Means by up to +6.0 pp and Budget-RAG by +1.9 pp.

Consolidator attaches a shared slot-local Consolidator, hierarchical router, routed short-term memory, persistent long-term memory, and sliding-attention KV cache to PMNet to learn replay-free STM-to-LTM consolidation. On the two-segment modulo-10 same-address update task, Consolidator reaches 87.02% updated-mapping LTM recall versus 18.32% for forced identity accumulation using only 12.35K trainable parameters.

ContextPilot augments Perception & Planning, Information Retrieval, Memory Management, and Context Offloading tools, then trains them with context-aware partial rollout and fine-grained credit assignment. On long-context QA, ContextPilot-14B-RL scores 72.20 average across NovelQA, ∞Bench, LongMemEval-S, and BrowseComp+ compared to 70.11 for StateLM-14B-RL.

ContextWeaver organizes agent histories using Dependency-Aware Context Construction, Dependency Summarizer, and a Validation and Test Layer into a dependency graph that drives selective context weaving. On SWE-Bench Verified, ContextWeaver with Claude Sonnet 4 achieves 66.0% pass@1 versus 63.2% for the Sliding Window baseline under the same window size.

ContraMem builds Function Cards, Skill Cards, a Reflector, and a Curator that turn heterogeneous same-task trajectories into compact procedural memory. On GAIA2/ARE, ContraMem raises the source-target macro success rate from 26.2% with no memory to 55.3%, 30.0 points above Agent Workflow Memory.

ConvMemory v2 adds a recall-preserving cascade on top of ConvMemory v1, using a fine-tuned ms-marco-MiniLM-L-6-v2 cross-encoder, a strict anti-shortcut inference contract, and a token-evidence ablation to reorder only v1’s protected top-10 candidates. On the LoCoMo benchmark, ConvMemory v2 raises FULL MRR from 0.5824 to 0.6560 over ConvMemory v1 and comes within 0.013 MRR of mxbai-rerank-large-v1 while running about 68× cheaper than the full-pool cross-encoder.

CoreMem combines Riemannian retrieval, Fisher-guided discrete token distillation, Residual Metric Fusion, and an Edge-Cloud Hybrid Architecture to retrieve and compress long-term dialogue memories under strict VRAM budgets. On LOCOMO with MiniLM-L6, CoreMem-Fusion reaches Judge accuracy 0.540 versus 0.531 for NaiveRAG, with +4.51 pp Open-domain and +4.17 pp Temporal gains.

COUNTERMEM combines Module I: Propose Alternatives, Module II: Check Alternatives and Store Corrections, and Module III: Retrieve and Select Records to turn failed actions into reusable, verified counterfactual memory. On the four-domain suite with gpt-oss-120b and ReAct, COUNTERMEM reaches a 57.5 average score versus 39.0 for ReAct without augmentation (+18.5 points) and reduces task-run tokens from 6.0M to 4.4M.

CreaMem organizes long-term agent memory into Meta Memory Manager, Episodic Memory, three Life Scene Memories, and a Core Memory, with a Planner LLM orchestrating per-memory balanced retrieval. On LongMemEval-S, CreaMem achieves 66.40% 4o-Judge accuracy versus 60.20% for MemGAS, and on LoCoMo reaches 54.61% versus 45.62% for HippoRAG 2.

CrystalMem combines Monitor, Crystallize, and Recrystallize around a four-state fidelity ladder, driven by advantage-weighted influence with dependency coupling and guarded by a verification gate. CrystalMem reaches 57.4 average restored capability across seven environments, beating R3Mem at 52.8 (+4.6 pp) under the same elastic budget.

CUE-Mem organizes synthetic multimodal user histories through Script Synthesis, Multimodal Dialog Synthesis, Data Review, and four tasks: Entity Recall, Long Pattern, Personalized Recommendation, and Answer Refusal. On CUE-Mem, textualized memory systems reach implicit Memory Avg. only 38.5 vs 80.8 with Oracle Evidence, exposing a large preservation and retrieval bottleneck for implicit cues.

CueMem extracts Memory Cue Extraction, builds a Turn Graph Construction, and runs Context Reconstruction plus Memory Management to answer long term dialogue questions. On LoCoMo, CueMem attains 84.07% accuracy, beating MemoryOS at 78.81% by 5.26 percentage points.

Diagnosing Retrieval vs. Utilization Bottlenecks in LLM Agent Memory combines Write Strategies, Retrieval Methods, and a Probing Framework to dissect memory-augmented agents. On LoCoMo, hybrid reranking reaches 77.2% accuracy versus 57.1% for BM25, proving retrieval choice dominates performance.

D-Mem combines Mem0∗, Quality Gating, and Full Deliberation into a dual-process memory system that incrementally stores vector memories and selectively scans raw history. On LoCoMo with GPT-4o-mini, D-Mem’s Quality Gating reaches 53.5 F1 versus the Mem0∗ baseline’s 51.2 F1, recovering 96.7% of the 55.3 F1 Full Deliberation performance with far fewer tokens.

Does Your Agent’s Memory Survive a Model Upgrade? decomposes agent memory into LC-RAW, RAG, NOTES, and KG-fixed with separate writer, reader, and embedder roles. On 48 synthetic histories, KG-fixed changes by only +0.0004 ± 0.0020 accuracy after a writer swap, while NOTES can shift by +9.91 or −13.28 percentage points and a 50/50 mixed embedding index forfeits most of an 11.90-point re-embedding gain.

DolphinBench builds Personas, History Planning, History Generation, Test Construction, and Test Verification into a hierarchical pipeline to create long-horizon, action-based memory tasks. On DolphinBench’s 600 tasks, the Hermes + GPT-5.6-Luna + Mem0 configuration reaches 70.67% accuracy while exposing distinct cost and latency tradeoffs against built-in memory and other baselines.

DP-MemView uses Slot-Level View Selection, Attribute Accounting, and the DP-MemView Interface Contract to turn raw memories into differentially private public views for response LLMs. On the PairedMem benchmark with Qwen2.5-7B, DP-MemView (on) achieves AUC 0.490 vs 0.842 for RawReadSet while maintaining U 0.877.