{
  "date": "2026-09-29",
  "domain": "Group-B3-Tech",
  "topic": "ai_agent_ecosystem",
  "evidence_items": [
    {
      "id": "EVD-TECH-20260929-01",
      "title": "E-mem: Multi-Agent Based Episodic Context Reconstruction for LLM Agent Memory",
      "type": "academic_paper",
      "url": "https://arxiv.org/abs/2601.21714",
      "publication_date": "2026-01-29",
      "authors": ["Kaixiang Wang", "Yidan Lin", "Zihan Wang", "Bunyod Suvonov", "Zhaojiacheng Zhou", "Yuxiang Zheng", "Jiaxi Cao", "Zhiheng Dong", "Chentao Wu", "Jiong Lou", "Jie Li"],
      "apa_citation": "Wang, K., Lin, Y., Wang, Z., Suvonov, B., Zhou, Z., Zheng, Y., ... & Li, J. (2026). E-mem: Multi-Agent Based Episodic Context Reconstruction for LLM Agent Memory. arXiv preprint arXiv:2601.21714. https://doi.org/10.48550/arXiv.2601.21714",
      "summary": "Proposes E-mem, a framework shifting from Memory Preprocessing to Episodic Context Reconstruction. Inspired by biological engrams, it employs a heterogeneous hierarchical architecture where multiple assistant agents maintain uncompressed memory contexts, while a central master agent orchestrates global planning. Resolves destructive de-contextualization in System 2 reasoning.",
      "pros": [
        "Prevents destructive de-contextualization typical in flat, pre-processed memory architectures.",
        "Hierarchical multi-agent design unloads raw context management to assistant agents.",
        "Facilitates high-precision System 2 reasoning over extended horizons."
      ],
      "cons": [
        "Keeping uncompressed context limits overall scalability due to high token consumption.",
        "Requires intricate synchronization between the master orchestrator and assistant agents."
      ],
      "confidence": "High"
    },
    {
      "id": "EVD-TECH-20260929-02",
      "title": "LEGOMem: Modular Procedural Memory for Multi-agent LLM Systems for Workflow Automation",
      "type": "academic_paper",
      "url": "https://arxiv.org/abs/2510.04851",
      "publication_date": "2025-10-06",
      "authors": ["Dongge Han", "Frederick Hoffman", "et al."],
      "apa_citation": "Han, D., Hoffman, F., et al. (2025). LEGOMem: Modular Procedural Memory for Multi-agent LLM Systems for Workflow Automation. arXiv preprint arXiv:2510.04851. https://doi.org/10.48550/arXiv.2510.04851",
      "summary": "Introduces a modular procedural memory framework that decomposes past task trajectories into reusable memory units. These units are flexibly allocated across orchestrators and task agents, improving planning, execution precision, and reducing execution failure rates by up to 16.2% on OfficeBench.",
      "pros": [
        "Provides modular execution-level guidance without introducing extra reasoning steps.",
        "Lowers failure rates and increases execution reliability.",
        "Compatible with various LLM team topologies (LLM-only, hybrid)."
      ],
      "cons": [
        "Depends on high-quality offline memory curation to generate procedural 'legos'.",
        "Potential context pollution if incorrect procedural units are retrieved."
      ],
      "confidence": "High"
    },
    {
      "id": "EVD-TECH-20260929-03",
      "title": "MIRIX: Multi-Agent Memory System for LLM-Based Agents",
      "type": "academic_paper",
      "url": "https://arxiv.org/abs/2507.07957",
      "publication_date": "2025-07-10",
      "authors": ["Yu Wang", "Xi Chen", "Julian McAuley"],
      "apa_citation": "Wang, Y., Chen, X., & McAuley, J. (2025). MIRIX: Multi-Agent Memory System for LLM-Based Agents. arXiv preprint arXiv:2507.07957. https://doi.org/10.48550/arXiv.2507.07957",
      "summary": "Develops MIRIX, consisting of six distinct memory types (Core, Episodic, Semantic, Procedural, Resource Memory, Knowledge Vault) coordinated by a multi-agent framework. Dynamically controls updates and retrieval via a routing mechanism, redefining long-term personalization and abstract recall.",
      "pros": [
        "Comprehensive taxonomy of memory (6 modules) covering virtually all agent needs.",
        "Agent-driven routing for updates and retrieval reduces monolithic prompt bottlenecks.",
        "Establishes a consensus-driven approach for scaling personalized AI persistence."
      ],
      "cons": [
        "Significant system complexity with six separate memory modules and routing logic.",
        "Heavyweight design might not be suitable for low-latency or lightweight edge tasks."
      ],
      "confidence": "High"
    },
    {
      "id": "EVD-TECH-20260929-04",
      "title": "Multi-Agent Deep Research: Training Multi-Agent Systems with M-GRPO",
      "type": "academic_paper",
      "url": "https://arxiv.org/abs/2511.13288",
      "publication_date": "2025-11-17",
      "authors": ["Haoyang Hong", "Jiajun Yin", "Yuan Wang", "Jingnan Liu", "Zhe Chen", "Ailing Yu", "Ji Li", "Zhiling Ye", "Hansong Xiao", "Yefei Chen", "Hualei Zhou", "Yun Yue", "Minghui Yang", "Chunxiao Guo", "Junwei Liu", "Peng Wei", "Jinjie Gu"],
      "apa_citation": "Hong, H., Yin, J., Wang, Y., Liu, J., Chen, Z., Yu, A., ... & Gu, J. (2025). Multi-Agent Deep Research: Training Multi-Agent Systems with M-GRPO. arXiv preprint arXiv:2511.13288. https://doi.org/10.48550/arXiv.2511.13288",
      "summary": "Proposes Multi-Agent Group Relative Policy Optimization (M-GRPO) to natively train multi-agent systems. Instead of a monolithic LLM, this approach fine-tunes distinct, specialized LLMs within a hierarchical multi-agent structure, overcoming optimization challenges caused by divergent role distributions.",
      "pros": [
        "Specializes models natively for specific agent roles, boosting deep research accuracy.",
        "M-GRPO resolves optimization instability in vertical multi-agent systems.",
        "Significantly outperforms unified-model baselines in specialized reasoning tasks."
      ],
      "cons": [
        "Requires maintaining and serving multiple fine-tuned models simultaneously.",
        "Training pipeline requires complex multi-agent reward assignment."
      ],
      "confidence": "High"
    },
    {
      "id": "EVD-TECH-20260929-05",
      "title": "Deep Research Agents: A Systematic Examination And Roadmap",
      "type": "academic_paper",
      "url": "https://arxiv.org/abs/2506.18096",
      "publication_date": "2025-06-22",
      "authors": ["Yuxuan Huang", "et al."],
      "apa_citation": "Huang, Y., et al. (2025). Deep Research Agents: A Systematic Examination and Roadmap. arXiv preprint arXiv:2506.18096. https://doi.org/10.48550/arXiv.2506.18096",
      "summary": "A systematic taxonomy of Deep Research Agents, covering dynamic reasoning, adaptive long-horizon planning, multi-hop retrieval, and structured reporting. Differentiates between static and dynamic workflows and sets evaluation roadmaps.",
      "pros": [
        "Offers a comprehensive architectural breakdown of modern deep research agents.",
        "Provides clear evaluation frameworks for multi-hop retrieval and planning.",
        "Highlights the shift from static pipelines to dynamic, self-correcting workflows."
      ],
      "cons": [
        "Being a survey/roadmap, it provides conceptual frameworks rather than novel algorithmic breakthroughs.",
        "Relies on existing benchmarks which may not fully capture long-horizon nuances."
      ],
      "confidence": "Medium"
    },
    {
      "id": "EVD-TECH-20260929-06",
      "title": "Microsoft LEGOMem (frederickhoffman/legomem)",
      "type": "open_source_repository",
      "url": "https://github.com/frederickhoffman/legomem",
      "authors": ["Frederick Hoffman", "Dongge Han", "Microsoft Research"],
      "graft_analysis": {
        "files": 21,
        "symbols": 52,
        "edges": 156,
        "tokens_saved": 10668,
        "core_components": ["legomem/core/orchestrator.py", "legomem/memory/vector_store.py", "legomem/eval/evaluator.py", "legomem/bench/datasets.py"]
      },
      "summary": "Reference implementation for modular procedural memory in multi-agent LLM systems. Evaluates workflows via LangSmith and weights-and-biases tracking.",
      "pros": [
        "Clear architectural separation between Core (Orchestrator, TaskAgent) and Memory (VectorStore).",
        "Includes end-to-end evaluation and W&B monitoring integration.",
        "Graft mapping saves 10k+ tokens and highlights 'create_legomem_graph' as a key hotspot."
      ],
      "cons": [
        "Heavily tied to LangSmith and OpenAI dependencies out of the box.",
        "Lacks built-in multi-modal memory capabilities."
      ],
      "confidence": "High"
    }
  ]
}