{
  "version": 1,
  "article": {
    "path": "evomembench-agent-memory-budget-replication",
    "title": "16K 到 128K，Agent 记忆增益从 +18 缩至 +6：EvoMemBench 子集复测与 Token 成本解剖",
    "locale": "zh-cn",
    "canonical": "https://vxi.me/posts/evomembench-agent-memory-budget-replication/",
    "markdown": "https://vxi.me/posts/evomembench-agent-memory-budget-replication.md"
  },
  "verification": {
    "level": "first-party",
    "verified_at": "2026-11-02T00:00:00.000Z",
    "summary": "预注册子集复测：Finding 3 的增益收窄方向复现（最佳增益 +18.0 -> +6.0 分），128K 跌破基线子句在本子集未复现，复合判读 PARTIAL；ACE 记忆模块的 token 成本为我方一手测量，论文未报告。",
    "environments": [
      "官方仓库 DSAIL-Memory/EvoMemBench @5f550e0c，InEp-Exec 子集（BFCL multi_turn_ours）官方 runner 与评分",
      "Gorilla File System 环境，仓库实发的全部 50 条样本，三臂 × 两档预算共 300 个 episode",
      "骨干为第三方网关标注为 gpt-6-luna 的端点（标注身份无法独立核验，判读以该标注端点为界）"
    ],
    "evidence_mode": "mixed"
  },
  "public_sources": {
    "repository": {
      "url": "https://github.com/vxime/evomembench-agent-memory-replication",
      "commit": "7e31af44c93fa88fb130e9417d36951b5ca1ac9d",
      "label": "复现回执、逐题数据与分析脚本"
    },
    "run": null
  }
}