{ "manifest_version": "0.2.0", "verified_at": "2026-09-20", "targets": [ { "id": "factual_recall", "label": "Factual recall", "description": "Recover information stated in prior interactions." }, { "id": "multi_session_reasoning", "label": "Multi-session reasoning", "description": "Combine evidence distributed across sessions." }, { "id": "temporal_reasoning", "label": "Temporal state & updates", "description": "Reason about time, changing facts, and historical state." }, { "id": "contradiction_preservation", "label": "Contradiction preservation", "description": "Keep conflicting or superseded evidence inspectable." }, { "id": "abstraction", "label": "Abstraction & consolidation", "description": "Control detail level or consolidate memory without erasing evidence boundaries." }, { "id": "episodic_planning", "label": "Episodic planning", "description": "Use prior experience to choose actions in an environment." }, { "id": "end_task_behavior", "label": "End-task behavior", "description": "Measure whether memory improves the downstream task, not only retrieval." }, { "id": "abstention", "label": "Abstention & missing evidence", "description": "Recognize when the available memory cannot support an answer." } ], "artifacts": [ { "id": "paper:2603.08965", "title": "Semantic Level of Detail for Knowledge Graphs: Discovering Abstraction Boundaries via Spectral Heat Diffusion", "artifact_type": "method", "primary_url": "https://huggingface.co/papers/2603.08965", "authors": ["Edward Izgorodin"], "mnemoverse_authored": true, "summary": "A continuous graph-zoom method that detects qualitative abstraction boundaries through spectral heat diffusion.", "memory_scope": ["knowledge graph", "hierarchical representation"], "evaluation_targets": ["abstraction"], "memory_unit": "Graph nodes and hierarchical communities", "time_model": "unknown", "conflict_update_model": "unknown", "outcome_level": "unknown", "evidence_status": "partially_evaluated", "code_url": "https://github.com/mnemoverse/mnemoverse-slod-paper", "project_url": null, "dataset_url": null, "license": "Apache-2.0 (code repository; see LICENSE and NOTICE)", "limitations": ["Evaluations cover abstraction boundaries on planted HSBM hierarchies and WordNet, not memory-assisted QA or agent-task completion.", "Choosing an abstraction scale from a user's query and transferring the defaults to implicit or different hierarchies remain open questions.", "Reported evaluations are author-reported; independent replication is not established by this source review."], "sources": ["https://huggingface.co/papers/2603.08965", "https://huggingface.co/api/papers/2603.08965", "https://arxiv.org/html/2603.08965v2", "https://github.com/mnemoverse/mnemoverse-slod-paper", "https://github.com/mnemoverse/mnemoverse-slod-paper/blob/062a68876d4eb3ffdeaedb1459f2eb4e9955434e/LICENSE", "https://github.com/mnemoverse/mnemoverse-slod-paper/blob/062a68876d4eb3ffdeaedb1459f2eb4e9955434e/NOTICE"], "verified_at": "2026-09-20" }, { "id": "paper:2410.10813", "title": "LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory", "artifact_type": "benchmark", "primary_url": "https://huggingface.co/papers/2410.10813", "authors": ["Di Wu", "Hongwei Wang", "Wenhao Yu", "Yuwei Zhang", "Kai-Wei Chang", "Dong Yu"], "mnemoverse_authored": false, "summary": "A 500-question benchmark for five long-term interactive-memory abilities in chat assistants.", "memory_scope": ["chat history", "multi-session memory"], "evaluation_targets": ["factual_recall", "multi_session_reasoning", "temporal_reasoning", "abstention", "end_task_behavior"], "memory_unit": "Sessions, conversation rounds, summaries and extracted user facts, evaluated as alternative representations", "time_model": "Temporal questions over scalable multi-session histories", "conflict_update_model": "Knowledge-update questions are an explicit benchmark ability", "outcome_level": "mixed", "evidence_status": "evaluated", "code_url": "https://github.com/xiaowu0162/LongMemEval", "project_url": null, "dataset_url": "https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned", "license": "MIT (code repository; cleaned dataset card also declares MIT)", "limitations": ["Recall@k and NDCG@k are reported separately from LLM-judged QA; agent-task completion is not evaluated.", "In the reported error analysis, some knowledge-update cases have correct answers but fail the strict retrieval criterion because only updated evidence, not earlier information, was retrieved.", "Reported outcomes remain specific to the benchmark's constructed histories and evaluation setup.", "The linked dataset is the authors' cleaned replacement, not a byte-identical archive of the original evaluation data."], "sources": ["https://huggingface.co/papers/2410.10813", "https://huggingface.co/api/papers/2410.10813", "https://arxiv.org/html/2410.10813v2", "https://github.com/xiaowu0162/LongMemEval", "https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned", "https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/blob/98d7416c24c778c2fee6e6f3006e7a073259d48f/README.md"], "verified_at": "2026-09-20" }, { "id": "paper:2501.13956", "title": "Zep: A Temporal Knowledge Graph Architecture for Agent Memory", "artifact_type": "framework", "primary_url": "https://huggingface.co/papers/2501.13956", "authors": ["Preston Rasmussen", "Pavlo Paliychuk", "Travis Beauvais", "Jack Ryan", "Daniel Chalef"], "mnemoverse_authored": false, "summary": "A temporal knowledge-graph memory layer evaluated with Deep Memory Retrieval and LongMemEval.", "memory_scope": ["conversation", "business data", "temporal knowledge graph"], "evaluation_targets": ["factual_recall", "multi_session_reasoning", "temporal_reasoning", "end_task_behavior"], "memory_unit": "Entities, facts, episodes, and historical relationships", "time_model": "Four edge timestamps distinguish system creation/invalidation from the period when a fact is valid.", "conflict_update_model": "An LLM compares new and related existing edges for temporally overlapping contradictions; affected edges are invalidated at the new edge's valid time, with ingestion order taking priority.", "outcome_level": "answer", "evidence_status": "evaluated", "code_url": "https://github.com/getzep/graphiti", "project_url": null, "dataset_url": null, "license": "Apache-2.0 (code repository)", "limitations": ["Reported evaluations are author-reported; independent replication is not established by this source review.", "The experiments measure answer quality, latency and context size, not retrieval-only quality or agent-task completion.", "Historical relationships and source links are described, but lossless history reconstruction and stale-source reimport are not separately evaluated."], "sources": ["https://huggingface.co/papers/2501.13956", "https://huggingface.co/api/papers/2501.13956", "https://arxiv.org/html/2501.13956v1", "https://github.com/getzep/graphiti"], "verified_at": "2026-09-20" }, { "id": "paper:2407.04363", "title": "AriGraph: Learning Knowledge Graph World Models with Episodic Memory for LLM Agents", "artifact_type": "framework", "primary_url": "https://huggingface.co/papers/2407.04363", "authors": ["Petr Anokhin", "Nikita Semenov", "Artyom Sorokin", "Dmitry Evseev", "Andrey Kravchenko", "Mikhail Burtsev", "Evgeny Burnaev"], "metadata_source": "https://arxiv.org/abs/2407.04363v3", "mnemoverse_authored": false, "summary": "A semantic and episodic graph memory evaluated in TextWorld and NetHack, and on static multi-hop question answering.", "memory_scope": ["semantic memory", "episodic memory", "environment model"], "evaluation_targets": ["factual_recall", "episodic_planning", "end_task_behavior"], "memory_unit": "Interconnected concepts and episodes acquired during environment interaction", "time_model": "Interaction episodes accumulate during environment exploration", "conflict_update_model": "Triplets from new observations are compared with existing semantic edges; outdated edges are removed. Preservation and historical retrieval of conflicting versions remain unverified.", "outcome_level": "mixed", "evidence_status": "evaluated", "code_url": "https://github.com/AIRI-Institute/AriGraph", "project_url": null, "dataset_url": null, "license": "MIT (code repository)", "limitations": ["Reported evaluations are author-reported; independent replication is not established by this source review.", "TextWorld averages the best three of five attempts; NetHack reports score and depth over three runs, not full-game completion.", "Static QA uses 200 sampled examples each from MuSiQue and HotpotQA; contradiction preservation remains unverified."], "sources": ["https://huggingface.co/papers/2407.04363", "https://huggingface.co/api/papers/2407.04363", "https://arxiv.org/abs/2407.04363v3", "https://arxiv.org/html/2407.04363v3", "https://github.com/AIRI-Institute/AriGraph"], "verified_at": "2026-09-20" }, { "id": "paper:2507.03724", "title": "MemOS: A Memory OS for AI System", "artifact_type": "framework", "primary_url": "https://huggingface.co/papers/2507.03724", "authors": ["Zhiyu Li", "Chenyang Xi", "Chunyu Li", "Ding Chen", "Boyu Chen", "Shichao Song", "Simin Niu", "Hanyu Wang", "Jiawei Yang", "Chen Tang", "Qingchen Yu", "Jihao Zhao", "Yezhaohui Wang", "Peng Liu", "Zehao Lin", "Pengyuan Wang", "Jiahao Huo", "Tianyi Chen", "Kai Chen", "Kehang Li", "Zhen Tao", "Huayi Lai", "Hao Wu", "Bo Tang", "Zhengren Wang", "Zhaoxin Fan", "Ningyu Zhang", "Linfeng Zhang", "Junchi Yan", "Mingchuan Yang", "Tong Xu", "Wei Xu", "Huajun Chen", "Haofen Wang", "Hongkang Yang", "Wentao Zhang", "Zhi-Qin John Xu", "Siheng Chen", "Feiyu Xiong"], "metadata_source": "https://arxiv.org/abs/2507.03724v4", "mnemoverse_authored": false, "summary": "A memory operating-system framework evaluated on answer quality and personalization using LoCoMo, LongMemEval, PreFEval, and PersonaMem.", "memory_scope": ["plaintext memory", "activation memory", "parameter memory"], "evaluation_targets": ["factual_recall", "multi_session_reasoning", "temporal_reasoning", "abstraction", "end_task_behavior"], "memory_unit": "MemCube with content, provenance, versioning, and metadata", "time_model": "Memory types can evolve, migrate, compose, and fuse over time", "conflict_update_model": "Knowledge updates are tested within LongMemEval; preservation of superseded versions remains unverified", "outcome_level": "answer", "evidence_status": "evaluated", "code_url": "https://github.com/MemTensor/MemOS", "project_url": "https://memos.openmem.net/", "dataset_url": null, "license": "Apache-2.0 (code repository)", "limitations": ["Reported evaluations are author-reported; independent replication is not established by this source review.", "Answer quality and preference alignment are evaluated separately from retrieval/API latency and KV acceleration; API success is not an agent-task completion rate.", "Preservation of conflicting versions and a shared valid-time/transaction-time contract remain unverified."], "sources": ["https://huggingface.co/papers/2507.03724", "https://huggingface.co/api/papers/2507.03724", "https://arxiv.org/abs/2507.03724v4", "https://arxiv.org/html/2507.03724v4", "https://github.com/MemTensor/MemOS", "https://memos.openmem.net/"], "verified_at": "2026-09-19" }, { "id": "paper:2602.05665", "title": "Graph-based Agent Memory: Taxonomy, Techniques, and Applications", "artifact_type": "survey", "primary_url": "https://huggingface.co/papers/2602.05665", "authors": ["Chang Yang", "Chuang Zhou", "Yilin Xiao", "Su Dong", "Luyao Zhuang", "Yujing Zhang", "Zhu Wang", "Zijin Hong", "Zheng Yuan", "Zhishang Xiang", "Shengyuan Chen", "Huachi Zhou", "Qinggang Zhang", "Ninghao Liu", "Jinsong Su", "Xinrun Wang", "Yi Chang", "Xiao Huang"], "mnemoverse_authored": false, "summary": "A survey of graph-based agent memory taxonomy, lifecycle techniques, open-source libraries, benchmarks, and applications.", "memory_scope": ["short- and long-term memory", "knowledge and experience memory", "graph-based memory"], "evaluation_targets": ["abstraction", "end_task_behavior"], "memory_unit": "Taxonomy spans non-structural and structural memory representations", "time_model": "Memory evolution is one stage of the surveyed lifecycle", "conflict_update_model": "The survey covers memory evolution; one common evaluated conflict model is not asserted", "outcome_level": "unknown", "evidence_status": "not_evaluated", "code_url": null, "project_url": "https://github.com/DEEP-PolyU/Awesome-GraphMemory", "dataset_url": null, "license": "unknown", "limitations": ["This is a survey and resource map, not a benchmark result.", "Coverage should not be interpreted as validation of every listed system or dataset."], "sources": ["https://huggingface.co/papers/2602.05665", "https://huggingface.co/api/papers/2602.05665", "https://arxiv.org/html/2602.05665v1", "https://github.com/DEEP-PolyU/Awesome-GraphMemory"], "verified_at": "2026-09-20" }, { "id": "paper:2608.15008", "title": "Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents", "artifact_type": "paper", "primary_url": "https://huggingface.co/papers/2608.15008", "authors": ["Wei-Chieh Huang", "Weizhi Zhang", "Yuchen Wu", "Yankai Chen", "Eric Hanchen Jiang", "Wooseong Yang", "Yiwei Yang", "Henry Peng Zou", "Hanrong Zhang", "Ying Nian Wu", "Haolun Wu", "Kai-Wei Chang", "Philip S. Yu", "Xue Liu", "Aylin Caliskan"], "mnemoverse_authored": false, "summary": "Controlled experiments examine how memory-substrate trade-offs change with the task, retrieval depth, and history length.", "memory_scope": ["multi-session dialogues", "agent trajectories", "cross-task experience"], "evaluation_targets": ["factual_recall", "multi_session_reasoning", "temporal_reasoning", "episodic_planning", "end_task_behavior"], "memory_unit": "Multiple substrate families under a shared harness, not one common representation", "time_model": "Within- and cross-episode histories; no shared valid-time/transaction-time contract", "conflict_update_model": "MAB Conflict Resolution tests most-recent-fact selection, not preservation of both versions", "outcome_level": "mixed", "evidence_status": "evaluated", "code_url": null, "project_url": null, "dataset_url": null, "license": "unknown", "limitations": ["Reported evaluations are author-reported; independent replication is not established by this source review.", "MemGPT, Mem0, and Zep appear only in an auxiliary-cost analysis (Appendix B.3), not in the main controlled comparison; M8 and M11 differ from their reference implementations.", "Code and harness-data release are unverified; code is promised upon acceptance.", "The authors argue for adaptive routing; they do not demonstrate a released router or a production-vendor ranking."], "sources": ["https://huggingface.co/papers/2608.15008", "https://huggingface.co/api/papers/2608.15008", "https://arxiv.org/html/2608.15008v1"], "verified_at": "2026-09-19" } ] }