{
  "date": "2026-09-28",
  "domain": "Group-B3-Tech",
  "topic": "ai_agent_ecosystem",
  "evidence_items": [
    {
      "id": "EVD-TECH-20260928-01",
      "title": "Mind DeepResearch Technical Report",
      "type": "academic_paper",
      "url": "https://arxiv.org/abs/2604.14518",
      "publication_date": "2026-04-14",
      "authors": ["Sheng Yang", "Yi Yao", "He Zhu", "Piaohong Wang", "Jincheng Ren", "Wangchunshu Zhou"],
      "apa_citation": "Yang, S., Yao, Y., Zhu, H., Wang, P., Ren, J., & Zhou, W. (2026). Mind DeepResearch Technical Report. arXiv preprint arXiv:2604.14518. https://doi.org/10.48550/arXiv.2604.14518",
      "summary": "Efficient multi-agent deep research framework that achieves leading performance with ~30B models via a collaborative 3-agent architecture (Planning, DeepSearch, Report) and a 4-stage specialized training pipeline (SFT cold-start, Search-RL, Report-RL, and preference alignment). Introduces Extended Chain-of-Thought (XoT) memory across agent interactions.",
      "pros": [
        "Achieves state-of-the-art performance (51.8 on MindDR Bench, 52.5 on DeepResearch Bench) using cost-effective ~30B models.",
        "XoT memory preserves full provenance of retrieved information across multi-agent turns.",
        "Search-RL and Report-RL decouple retrieval optimization from long-form writing."
      ],
      "cons": [
        "Requires multi-stage RL training and complex reward modeling with LLM-as-judge.",
        "XoT shared memory can grow significantly during multi-hop deep search."
      ],
      "confidence": "High"
    },
    {
      "id": "EVD-TECH-20260928-02",
      "title": "Scaling Teams or Scaling Time? Memory Enabled Lifelong Learning in LLM Multi-Agent Systems",
      "type": "academic_paper",
      "url": "https://arxiv.org/abs/2604.03295",
      "publication_date": "2026-04-03",
      "authors": ["Shanglin Wu", "Lei Zhang", "Yuan Wang", "Xin Liu"],
      "apa_citation": "Wu, S., Zhang, L., Wang, Y., & Liu, X. (2026). Scaling teams or scaling time? Memory enabled lifelong learning in LLM multi-agent systems. arXiv preprint arXiv:2604.03295. https://doi.org/10.48550/arXiv.2604.03295",
      "code_url": "https://github.com/ShanglinWu/MAS_lifelong_learning",
      "summary": "Proposes LLMA-Mem, integrating episodic, procedural, and transactive memory ('who knows what') under flexible topologies (local, shared, hybrid). Demonstrates that larger teams introduce coordination overhead and that smaller teams with lifelong memory outperform larger teams while cutting token cost by 9.4% to 71.7%.",
      "pros": [
        "Empirically proves non-monotonic scaling where smaller teams with transactive memory beat large naive swarms.",
        "Reduces token costs substantially (up to 71.7%) across coding, research, and database benchmarks.",
        "Transactive memory structures agent profiles and team coordination patterns for optimal task allocation."
      ],
      "cons": [
        "Transactive memory requires maintenance of proficiency estimates and team composition models.",
        "Hybrid topologies require careful synchronization to avoid stale state."
      ],
      "confidence": "High"
    },
    {
      "id": "EVD-TECH-20260928-03",
      "title": "AdMem: Advanced Memory for Task-solving Agents",
      "type": "academic_paper",
      "url": "https://arxiv.org/abs/2606.06787",
      "publication_date": "2026-06-05",
      "authors": ["Runzhe Wang", "Huilin Lu", "Shengjie Liu", "Li Dong", "Jason Zhu"],
      "apa_citation": "Wang, R., Lu, H., Liu, S., Dong, L., & Zhu, J. (2026). AdMem: Advanced memory for task-solving agents. arXiv preprint arXiv:2606.06787. https://doi.org/10.48550/arXiv.2606.06787",
      "summary": "A bi-level memory framework integrating semantic, episodic, and procedural memory across short-term and long-term stores. Features an actor, memory, and critic multi-agent system with reward-based memory evaluation, merging, and pruning.",
      "pros": [
        "Addresses both failure cases and success trajectories in procedural memory.",
        "Bi-level design with reward-based pruning maintains constant memory footprint.",
        "Actor-Memory-Critic separation guarantees memory quality and prevents hallucination accumulation."
      ],
      "cons": [
        "Requires auxiliary LLM calls for critic evaluation and memory reorganization.",
        "Pruning policies must be tuned per domain to avoid removing critical historical edge cases."
      ],
      "confidence": "High"
    },
    {
      "id": "EVD-TECH-20260928-04",
      "title": "General Agentic Memory Via Deep Research",
      "type": "academic_paper",
      "url": "https://arxiv.org/abs/2511.18423",
      "publication_date": "2025-11-27",
      "authors": ["Jiamin Chen", "Shuo Wang", "Yifan Li", "Kun Zhao", "Min Zhang"],
      "apa_citation": "Chen, J., Wang, S., Li, Y., Zhao, K., & Zhang, M. (2025). General agentic memory via deep research. arXiv preprint arXiv:2511.18423. https://doi.org/10.48550/arXiv.2511.18423",
      "summary": "Proposes General Agentic Memory (GAM) based on Just-in-Time (JIT) compilation principles. Instead of lossy offline summarization, GAM retains complete historical data in a page-store with lightweight index (Memorizer), and uses an online Researcher agent to conduct iterative multi-hop deep research at query time.",
      "pros": [
        "Lossless retention of original historical context, outperforming ahead-of-time static memory on HotpotQA and RULER.",
        "Achieves >90% accuracy on multi-hop tracing tasks.",
        "Scales test-time compute flexibly via autonomous reflection depth (1 to 5 rounds)."
      ],
      "cons": [
        "Increases query-time latency due to on-demand deep research steps.",
        "Requires fast underlying database retrieval infrastructure (hybrid vector + BM25 + ID store)."
      ],
      "confidence": "High"
    },
    {
      "id": "EVD-TECH-20260928-05",
      "title": "DeepAgent: A General Reasoning Agent with Scalable Toolsets",
      "type": "academic_paper",
      "url": "https://arxiv.org/abs/2510.21618",
      "publication_date": "2025-10-24",
      "authors": ["Jiahao Xu", "Chao Wang", "Zhiwei Liu", "Xiao Chen"],
      "apa_citation": "Xu, J., Wang, C., Liu, Z., & Chen, X. (2025). DeepAgent: A general reasoning agent with scalable toolsets. arXiv preprint arXiv:2510.21618. https://doi.org/10.48550/arXiv.2510.21618",
      "code_url": "https://github.com/RUC-NLPIR/DeepAgent",
      "summary": "Introduces Autonomous Memory Folding (<fold_thought>) where an agent dynamically pauses, consolidates interaction history into structured JSON schema (episodic, working, tool memories) using an auxiliary LLM, and re-plans. Trained via Tool Policy Optimization (ToolPO) with fine-grained tool advantage attribution.",
      "pros": [
        "Memory folding allows agents to 'take a breath' and avoid catastrophic trajectory errors on GAIA and HLE.",
        "Structured JSON memory schema prevents loss of critical parameters compared to free-form text summarization.",
        "ToolPO provides stable RL credit assignment to individual tool call tokens."
      ],
      "cons": [
        "Auxiliary model invocation introduces overhead during memory folding triggers.",
        "Requires simulated API environments for effective ToolPO training."
      ],
      "confidence": "High"
    },
    {
      "id": "EVD-TECH-20260928-06",
      "title": "LangChain Open Deep Research (open_deep_research)",
      "type": "open_source_repository",
      "url": "https://github.com/langchain-ai/open_deep_research",
      "authors": ["LangChain Team"],
      "graft_analysis": {
        "files": 22,
        "symbols": 180,
        "edges": 485,
        "tokens_saved": 89013,
        "core_components": ["deep_researcher.py", "multi_agent.py", "utils.py", "auth.py", "evaluators.py"]
      },
      "summary": "Open-source autonomous deep research framework orchestrating supervisor and recursive research sub-agents with dynamic MCP integration, parallel search tools, and token overflow guardrails.",
      "pros": [
        "Production-grade recursive research graph with built-in token limit detection (is_token_limit_exceeded).",
        "Robust MCP tool loading and error recovery in exception chains (_find_mcp_error_in_exception_chain).",
        "Evaluated on 100 PhD-level research tasks using RACE score metric."
      ],
      "cons": [
        "Dependency on complex LangGraph state graph conventions.",
        "Legacy codebase components mixed with current open_deep_research module."
      ],
      "confidence": "High"
    }
  ]
}