{
  "schema_version": "1.0",
  "id": "s13:https://arxiv.org/abs/2607.24667v1",
  "slug": "2607-24667v1-177qf29",
  "url": "https://feed7.dev/p/2607-24667v1-177qf29",
  "title": "Eviction as Estimation: A Fixed-Lag Smoothing View of Test-Time Memory, and When Measuring Beats Accumulating",
  "why_included": "A delayed KV-eviction policy helps when memory reuse is sharp and observable, but loses its edge on natural-text benchmarks—a useful warning for agent-memory claims.",
  "summary": "RMM delays eviction long enough to observe which stored items support a correct near-future prediction. It is a training-free generalization of **H2O** and reduced to H2O when that utility measurement was uniform.",
  "practical_implication": "For builders tuning bounded agent memory, test eviction policies on the actual reuse pattern. RMM made small memory act much larger in controlled delayed-reuse settings, but was only **on par with H2O** for single-turn QA and trailed H2O and SnapKV in streaming multi-turn tests.",
  "agent_context": "RMM delays eviction long enough to observe which stored items support a correct near-future prediction. It is a training-free generalization of **H2O** and reduced to H2O when that utility measurement was uniform.\n\nFor builders tuning bounded agent memory, test eviction policies on the actual reuse pattern. RMM made small memory act much larger in controlled delayed-reuse settings, but was only **on par with H2O** for single-turn QA and trailed H2O and SnapKV in streaming multi-turn tests.\n\nOn natural text, correctness weighting changed attention too little, so demonstrated utility largely collapsed into accumulated attention. The proposed framework may help classify workloads, but the reported policy is not a new general-purpose state of the art.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2607.24667v1",
    "published_at": "2026-07-27T17:08:27.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "benchmark",
  "domains": [],
  "topics": [
    "agent-memory",
    "context-caching",
    "benchmark-integrity"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "On natural text, correctness weighting changed attention too little, so demonstrated utility largely collapsed into accumulated attention. The proposed framework may help classify workloads, but the reported policy is not a new general-purpose state of the art."
  ],
  "lifecycle": "Current",
  "published_at": "2026-07-27T17:08:27.000Z",
  "modified_at": "2026-07-27T17:08:27.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2607-24667v1-177qf29",
    "json": "https://feed7.dev/p/2607-24667v1-177qf29.json",
    "markdown": "https://feed7.dev/p/2607-24667v1-177qf29.md"
  }
}