Add weekly cache research report for 2026-05-24

[thinkingfish]

Summary

• Adds cache-research/weekly-cache-report-2026-05-24.md covering the 2026-05-18 → 2026-05-24 window (the seven days following the prior weekly).
• 10 primary entries — expanded from the ~5 target because MLSys 2026 ran during the window (May 18–22, Bellevue), bringing invited talks, posters, and vendor blog launches timed to the conference together with a Cohere model release and three new in-window arXiv KV-systems papers.

Entries

A. Production measurements and empirical studies

1. vLLM × Novita AI — PegaFlow external KV cache (Rust sidecar) — vLLM blog, May 18
2. Databricks — managed prompt caching, 2.5× throughput / 3× P50 on production GPT-OSS — May 22
3. MLSys 2026 — LMCache invited talk + Kitty 2-bit KV poster — Bellevue, May 18–22
4. Cohere — Command A+ launch with explicit 3:1 sliding-window/global attention ratio and full-precision KV path — May 20

B. Academic / idea-forward work

5. KVDrive — multi-tier KV management across HBM / DRAM / NVMe (arXiv:2605.18071, SIGMOD 2026)
6. KVServe — service-aware adaptive KV compression for disaggregated serving (arXiv:2605.13734)
7. PEEK — context map as an orientation cache for long-context LLM agents (arXiv:2605.19932)
8. GEM — GPU-variability-aware expert-to-GPU mapping for MoE (arXiv:2605.19945)
9. OScaR — rotation-based extreme KV quantization, multimodal (arXiv:2605.19660)
10. Protection Is (Nearly) All You Need — seven-policy eviction bake-off (arXiv:2605.18053)

Plus an Additional Context section (storage caching: Google Cloud Storage Rapid, MinIO MemKV; Red Hat / Snowflake adjacent reads; arXiv just-past-window items: TIDE, PulseCol, Runtime-Certified Quantized Attention) and cross-cutting observations (MLSys as a community event; external KV cache as a separate-process pattern; KV-cache footprint moving into pretraining; eviction-policy methodology catching up; the cache abstraction expanding beyond KV).

Test plan

• Source-sweep cross-check: frontier AI labs, inference vendors, storage vendors, systems/ML venue accepted lists, arXiv 2605.13xxx–2605.21xxx sweep, curated trackers, HN.
• Each entry has a primary reference linked in text and listed in the References section.
• Runtime-evaluation status noted explicitly for every entry (including "partial" / "no" cases).
• Explicit-negatives list covers every named source class.
• Methodology caveat documented (arxiv.org / mlsys.org / vllm.ai 403s on direct fetch; figures flagged as "as-claimed").

https://claude.ai/code/session_01EroHYQeTaDJnvUoDgcLbnG

---

Generated by Claude Code

[thinkingfish]

Moved to iopsystems/inference-systems#3, where the weekly cache research reports now live under research/. Closing here.

---

Generated by Claude Code