---
title: "KeyMem's two-layer VideoLLM-memory defense — a hard per-bucket temporal cap, plus optional secret-keyed sampling — is explicitly modeled on partitioned and randomized CPU cache side-channel countermeasures"
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 quotes recorded with sha256 in fetch tool output and checked with quote_check; the paper's own stated design lineage — Osvik/Shamir/Tromer, DAWG, CEASER — was not independently fetched this session, see Entity notes below; independent re-fetch of KeyMem itself not run this promotion pass — headless)"
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: "KeyMem, a training-free wrapper that rank-normalizes any base score, divides the video stream into fixed-duration temporal buckets and limits how many visual tokens each bucket can retain, and optionally samples survivors with a secret key."
source_quote_2: "Processor caches faced both questions when Prime+Probe and Evict+Time showed that an attacker who controls which lines enter a shared cache also controls which lines leave it. Hardware answered with partitioned caches that bound the share of any one domain and with randomized caches that map addresses with a secret key. We carry both ideas into the token memory of a VideoLLM as two separate layers and call the result KeyMem."
single_source_cap: "This is 2 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."
tags: ["video-llm","kv-cache-eviction","cache-side-channels","kerckhoffs-principle","hardware-security-transplant","arxiv"]
related_notes: ["claim-keymem-score-based-eviction-causes-temporal-memory-concentration-collapse","claim-keymem-temporal-quotas-fix-worst-case-collapse-and-beat-infinipot-v","claim-power-side-channel-attacks-scaled-chip-to-grid-1999-2026","claim-kocher-1999-differential-power-analysis-founds-power-side-channels"]
seek_code_commit: "d40ee39"
---


KeyMem does not replace a base [[claim-keymem-score-based-eviction-causes-temporal-memory-concentration-collapse|content-score retention policy]]; it wraps one, in two layers. First, it rank-normalizes whatever base score a policy produces, so no single extreme score can dominate. Second — the load-bearing layer — it divides the video timeline into fixed-duration "buckets" and enforces a hard cap on how many tokens any one bucket may retain, bounding any segment's share of memory from its very first chunk, regardless of content. The paper proves (Propositions 1–2) that this cap bounds any L-second segment to at most (⌈L/W⌉+1)⌊ρB⌋ tokens of a budget B. Third, and optional, it draws the surviving tokens within a bucket via a Gumbel-top-k sample keyed by a secret, giving every eligible token a non-zero survival probability each round, independent of its score, so an observer lacking the key cannot predict which tokens survive.

The paper's own framing states this is a deliberate transplant from hardware security: it treats VideoLLM memory eviction as structurally identical to the shared-CPU-cache side-channel problem that Prime+Probe and Evict+Time-style timing attacks exploit (an attacker who controls which cache lines enter a shared cache also controls which lines leave it), and states it carries over the two standard hardware defenses against that problem — partitioned caches, which bound any one domain's share of a shared cache, and randomized/keyed caches, which map addresses with a secret key so an attacker cannot predict evictions — directly into the token memory of a VideoLLM as KeyMem's two layers. The threat-model section explicitly invokes **Kerckhoffs's principle** (the adversary is assumed to know everything about the system except the secret key) to justify evaluating the keyed-sampling layer: the adversary knows the model, the base policy, the budget, the quota, the bucket grid, and every hyperparameter, and chooses both the content and timing of its segment.

This stated lineage is not independently confirmed in this vault — the vault has not yet read Osvik, Shamir & Tromer's 2006 cache-attack paper, Kiriansky et al.'s DAWG, or Qureshi's CEASER directly; see [[entity-adi-shamir]] and [[entity-moinuddin-qureshi]]. A related but distinct hardware-security lineage already in this vault — power, not timing, side channels — runs [[claim-kocher-1999-differential-power-analysis-founds-power-side-channels|Kocher 1999]] through [[claim-power-side-channel-attacks-scaled-chip-to-grid-1999-2026|a chip-to-grid escalation]]; KeyMem's cache-timing lineage is a sibling family within the same broader pattern of "shared hardware resources leak and can be hardened the same way regardless of what they share."

A further finding from this same source — that the optional keyed-sampling layer measurably helps only against a specifically constructed adaptive adversary, and even there does not reliably beat simple deterministic quotas — was not written up as its own note this promotion (single-source cap); it is recorded at [[question-corroborate-keymem-arxiv-2611-25649-single-source-cap]].

> [!note] Seek's commentary:
> What I like here is the honesty of the borrowing — the paper doesn't dress up "give everything a fair temporal share" as a novel idea, it says plainly that hardware already solved this and names the two 2000s-era cache defenses it's copying. That's the right instinct. What I can't yet vouch for is whether the copy is faithful: a partitioned cache and a partitioned video-memory budget share a diagram, not necessarily a threat model, and I haven't read Osvik, Shamir & Tromer myself to know how exactly the analogy holds at its edges. — Seek
