---
title: "Imposing any temporal-allocation quota removes streaming-VideoLLM memory's worst-case accuracy collapse and outperforms the published InfiniPot-V baseline on two of three tested models"
type: "claim"
status: "seedling"
audit_status: "capture-verified (capture writer fetched the arXiv PDF directly via extract_pdf at capture time, 2026-11-21, exact quote recorded with sha256 in fetch tool output and checked with quote_check; InfiniPot-V's own published numbers were not independently fetched this session — this claim rests on KeyMem's own report of the comparison, not a reading of InfiniPot-V itself; see Entity notes below and [[question-corroborate-keymem-arxiv-2611-25649-single-source-cap]])"
writer_model: "claude-sonnet-5"
date_created: "2026-11-21T00:00:00.000Z"
provenance: "Promotion from 10-inbox/raw/2026-11-21-outside-arxiv261125649-keymem-temporal-occupancy-control-for-robust.md, 2026-11-21"
origin: "batch"
derived_from: ["10-inbox/raw/2026-11-21-outside-arxiv261125649-keymem-temporal-occupancy-control-for-robust.md"]
source_url: "https://arxiv.org/pdf/2611.25649"
source_sha: "3f9e83f2dc2afdc30540eea3864c201947827ba785f356504d34094129ddfa2f"
source_title: "KeyMem: Temporal Occupancy Control for Robust Streaming VideoLLM Memory"
source_author: "Wei Song, Yuxin Cao, Xiao Cheng, Christy Jie Liang, Xi Zheng, Leo Yu Zhang"
source_date: "2026-11-19T00:00:00.000Z"
source_venue: "arXiv:2611.25649v1 [cs.CY]"
source_tier: 1
source_quote: "Across five natural stressors, three VideoLLMs and two budgets, temporal quotas raise the worst accuracy of attention-based retention at the smaller budget from 42.2% to 73.3–76.1% and exceed that of InfiniPot-V by 6.1–16.7% on two of the three models."
single_source_cap: "This is 3 of 3 claim-notes resting on this single unrefereed arXiv preprint (2611.25649, posted 2026-11-19, no indication of peer review). Cap reached this promotion — see [[question-corroborate-keymem-arxiv-2611-25649-single-source-cap]] for the deferred fourth finding and what independent corroboration is needed, including an independent read of InfiniPot-V itself."
tags: ["video-llm","kv-cache-eviction","benchmark-comparison","robustness","arxiv"]
related_notes: ["claim-keymem-score-based-eviction-causes-temporal-memory-concentration-collapse","claim-keymem-two-layer-mechanism-modeled-on-cache-side-channel-defenses"]
seek_code_commit: "d40ee39"
---


Across five natural "stressors" (calm footage, rapid cuts, burned-in text, flicker, camera shake), three streaming VideoLLMs (Qwen2.5-VL-3B, Qwen2.5-VL-7B, LLaVA-OneVision-7B), and two memory budgets, KeyMem (arXiv:2611.25649) reports that adding *any* temporal-allocation rule — equal time-spans, a hard cap alone, or the full [[claim-keymem-two-layer-mechanism-modeled-on-cache-side-channel-defenses|KeyMem wrapper]] — removes the worst-case accuracy collapse that pure attention-based [[claim-keymem-score-based-eviction-causes-temporal-memory-concentration-collapse|content-score retention]] otherwise suffers on some models. At the smaller of the two tested memory budgets, the worst-case accuracy of attention-based retention rises from 42.2% to 73.3–76.1% once a temporal quota of any kind is added. These temporal-quota methods also exceed the previously published InfiniPot-V retention rule (Minsoo Kim, Kyuhong Shim, Jungwook Choi, Simyung Chang; NeurIPS 2026) by 6.1–16.7% on two of the three model families tested.

The comparison to InfiniPot-V is one-sided: this claim rests on KeyMem's own paper's report of how it compares to InfiniPot-V, not on an independent reading of InfiniPot-V's own published results. InfiniPot-V itself was not fetched in this session. That gap, and the open question of whether the improvement holds on the third model family, are routed to [[question-corroborate-keymem-arxiv-2611-25649-single-source-cap]] rather than left silently inside a clean-looking percentage.

> [!note] Seek's commentary:
> "Any temporal-allocation rule" doing most of the lifting here is a different claim than "KeyMem specifically is the answer" — equal time-spans, a hard cap, and the full keyed wrapper all cross the worst-case line together. That's a point in favor of the diagnosis (unbounded temporal concentration is the actual disease) over the cure (KeyMem's specific machinery), and it's worth remembering which part of this result is doing the work before crediting the fancier of the three fixes. — Seek
