{
  "title": "Memory, context management and selective observation in agents",
  "question": "What information is retained, when is it selected, and how is selection trained?",
  "scope": "Qualitative comparisons of methods, information requirements and task design in the cited paper versions.",
  "keywords": [
    "agent memory",
    "context management",
    "selective observation",
    "partial observability",
    "context budget"
  ],
  "dimensions": [
    "Information retained",
    "Operation and timing",
    "Training or selection rule",
    "Task setting"
  ],
  "html_url": "https://hoyant-su-agentic-rl.hf.space/topics/comparisons/agent-memory-observation.html",
  "json_url": "https://hoyant-su-agentic-rl.hf.space/topics/comparisons/agent-memory-observation.json",
  "csv_url": "https://hoyant-su-agentic-rl.hf.space/topics/comparisons/agent-memory-observation.csv",
  "bibtex_url": "https://hoyant-su-agentic-rl.hf.space/topics/comparisons/agent-memory-observation.bib",
  "paper_count": 4,
  "rows": [
    {
      "label": "MemAgent",
      "paper": {
        "name": "MemAgent",
        "version": "2507.02259v2",
        "title": "MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent",
        "authors": [
          "Yu, Hongli",
          "Chen, Tinghong",
          "Feng, Jiangtao",
          "Chen, Jiangjie",
          "Dai, Weinan",
          "Yu, Qiying",
          "Zhang, Ya-Qin",
          "Ma, Wei-Ying",
          "Liu, Jingjing",
          "Wang, Mingxuan",
          "Zhou, Hao"
        ],
        "publication_date": "2025/07/03",
        "paper_url": "https://arxiv.org/abs/2507.02259v2",
        "html_url": "https://arxiv.org/html/2507.02259v2"
      },
      "properties": {
        "Information retained": {
          "value": "Compact memory of previously read document chunks",
          "sources": [
            {
              "url": "https://arxiv.org/html/2507.02259v2#S2.SS1.p1",
              "section": "2.1 The MemAgent Workflow: RL-shaped Memory for Unbounded Contexts"
            }
          ]
        },
        "Operation and timing": {
          "value": "Overwrite memory after reading each chunk",
          "sources": [
            {
              "url": "https://arxiv.org/html/2507.02259v2#S2.SS1.p2",
              "section": "2.1 The MemAgent Workflow: RL-shaped Memory for Unbounded Contexts"
            }
          ]
        },
        "Training or selection rule": {
          "value": "Multi-conversation DAPO; final-answer advantage shared across conversations",
          "sources": [
            {
              "url": "https://arxiv.org/html/2507.02259v2#S2.SS2.p1",
              "section": "2.2 Training MemAgent with Multi-conv RL"
            },
            {
              "url": "https://arxiv.org/html/2507.02259v2#S2.SS2.p2",
              "section": "2.2 Training MemAgent with Multi-conv RL"
            }
          ]
        },
        "Task setting": {
          "value": "Document processing followed by answer generation",
          "sources": [
            {
              "url": "https://arxiv.org/html/2507.02259v2#S2.SS1.p3",
              "section": "2.1 The MemAgent Workflow: RL-shaped Memory for Unbounded Contexts"
            }
          ]
        }
      }
    },
    {
      "label": "AgentFold",
      "paper": {
        "name": "AgentFold",
        "version": "2510.24699v1",
        "title": "AgentFold: Long-Horizon Web Agents with Proactive Context Management",
        "authors": [
          "Ye, Rui",
          "Zhang, Zhongwang",
          "Li, Kuan",
          "Yin, Huifeng",
          "Tao, Zhengwei",
          "Zhao, Yida",
          "Su, Liangcai",
          "Zhang, Liwen",
          "Qiao, Zile",
          "Wang, Xinyu",
          "Xie, Pengjun",
          "Huang, Fei",
          "Chen, Siheng",
          "Zhou, Jingren",
          "Jiang, Yong"
        ],
        "publication_date": "2025/10/28",
        "paper_url": "https://arxiv.org/abs/2510.24699v1",
        "html_url": "https://arxiv.org/html/2510.24699v1"
      },
      "properties": {
        "Information retained": {
          "value": "Multiscale history summaries and the complete latest interaction",
          "sources": [
            {
              "url": "https://arxiv.org/html/2510.24699v1#S3.SS2.p2",
              "section": "3.2 AgentFold’s Context: Multi-Scale State Summaries, Latest Interaction"
            }
          ]
        },
        "Operation and timing": {
          "value": "Generate a folding directive alongside the next action",
          "sources": [
            {
              "url": "https://arxiv.org/html/2510.24699v1#S3.SS3.p1",
              "section": "3.3 AgentFold’s Response: Thinking, Folding, Explanation, Action"
            }
          ]
        },
        "Training or selection rule": {
          "value": "Supervised fine-tuning on generated and filtered trajectories",
          "sources": [
            {
              "url": "https://arxiv.org/html/2510.24699v1#S3.SS4.p1",
              "section": "3.4 AgentFold’s Training: Data Trajectory Collection"
            },
            {
              "url": "https://arxiv.org/html/2510.24699v1#S3.SS4.p2",
              "section": "3.4 AgentFold’s Training: Data Trajectory Collection"
            }
          ]
        },
        "Task setting": {
          "value": "Multi-step web information seeking",
          "sources": [
            {
              "url": "https://arxiv.org/html/2510.24699v1#S3.SS1.p1",
              "section": "3.1 Overview"
            }
          ]
        }
      }
    },
    {
      "label": "sigma-Reveal",
      "paper": {
        "name": "A3",
        "version": "2605.08013v1",
        "title": "Learning CLI Agents with Structured Action Credit under Selective Observation",
        "authors": [
          "Su, Haoyang",
          "Wen, Ying"
        ],
        "publication_date": "2026/05/08",
        "paper_url": "https://arxiv.org/abs/2605.08013v1",
        "html_url": "https://arxiv.org/html/2605.08013v1"
      },
      "properties": {
        "Information retained": {
          "value": "Task-relevant files and their directory context",
          "sources": [
            {
              "url": "https://arxiv.org/html/2605.08013v1#S3.SS2.p1",
              "section": "3.2 $\\sigma$ -Reveal Context Harness"
            }
          ]
        },
        "Operation and timing": {
          "value": "Select an initial workspace view before the first action",
          "sources": [
            {
              "url": "https://arxiv.org/html/2605.08013v1#S3.SS2.p1",
              "section": "3.2 $\\sigma$ -Reveal Context Harness"
            }
          ]
        },
        "Training or selection rule": {
          "value": "Relevance scoring and subtree-closed selection under a token budget",
          "sources": [
            {
              "url": "https://arxiv.org/html/2605.08013v1#S3.SS2.p1",
              "section": "3.2 $\\sigma$ -Reveal Context Harness"
            }
          ]
        },
        "Task setting": {
          "value": "Partially observed filesystem workspaces for CLI agents",
          "sources": [
            {
              "url": "https://arxiv.org/html/2605.08013v1#S3.SS2.p1",
              "section": "3.2 $\\sigma$ -Reveal Context Harness"
            }
          ]
        }
      }
    },
    {
      "label": "HiMPO",
      "paper": {
        "name": "HiMPO",
        "version": "2606.16285v2",
        "title": "HiMPO: Hindsight-Informed Memory Policy Optimization for Less-Entangled Credit in Long-Horizon Agents",
        "authors": [
          "Yan, Jiangze",
          "Shen, Yi",
          "Zhang, Wenjing",
          "Huang, Jieyun",
          "Liu, Zhaoxiang",
          "Wang, Ning",
          "Wang, Kai",
          "Lian, Shiguo"
        ],
        "publication_date": "2026/06/15",
        "paper_url": "https://arxiv.org/abs/2606.16285v2",
        "html_url": "https://arxiv.org/html/2606.16285v2"
      },
      "properties": {
        "Information retained": {
          "value": "Compressed memory with the latest interaction",
          "sources": [
            {
              "url": "https://arxiv.org/html/2606.16285v2#S3.SS1.p1",
              "section": "3.1 Problem Setup and Overview"
            }
          ]
        },
        "Operation and timing": {
          "value": "Write memory before reasoning and tool use at each step",
          "sources": [
            {
              "url": "https://arxiv.org/html/2606.16285v2#S3.SS1.p1",
              "section": "3.1 Problem Setup and Overview"
            }
          ]
        },
        "Training or selection rule": {
          "value": "Outcome-based RL plus memory-specific utility and hindsight credit",
          "sources": [
            {
              "url": "https://arxiv.org/html/2606.16285v2#S3.SS2.p2",
              "section": "3.2 Local Counterfactual Utility"
            },
            {
              "url": "https://arxiv.org/html/2606.16285v2#S3.SS3.p3",
              "section": "3.3 Retrospective Filtering of Memory Utility"
            },
            {
              "url": "https://arxiv.org/html/2606.16285v2#S3.SS4.p4",
              "section": "3.4 Stabilized Token-Level Policy Optimization"
            }
          ]
        },
        "Task setting": {
          "value": "Open-domain agent tasks and compressive-memory question answering",
          "sources": [
            {
              "url": "https://arxiv.org/html/2606.16285v2#S4.SS1.SSS0.Px1.p1",
              "section": "Benchmarks and metrics."
            }
          ]
        }
      }
    }
  ]
}