{"id":"razorattention--openreview-net","name":"RazorAttention","url":"https://openreview.net/forum?id=tkiZQlL04w","page":"https://mrpeppersdev.github.io/agent-infrastructure-landscape/systems/razorattention--openreview-net","family":"memory-research","section":"Recent method papers — theorized, no distinct product","sections":["Recent method papers — theorized, no distinct product"],"tier":3,"last_verified_at":"2026-10-23","lifecycle":{"status":"active","severity":0,"note":null},"taxonomy":{"storage":[{"value":"kv-cache","primary":true,"reason":null}],"retrieval":[{"value":"attention","primary":true,"reason":null},{"value":"cache-lookup","primary":false,"reason":null}],"persistence":[{"value":"session","primary":true,"reason":null}],"update":[{"value":"evict-oldest","primary":true,"reason":null}],"unit":[{"value":"kv-token","primary":true,"reason":null}],"governance":[{"value":"n/a","primary":true,"reason":"searched not found"}],"conflict":[{"value":"n/a","primary":true,"reason":"not applicable (KV-cache compression; no user write)"}]},"cells":{"type":{"value":"KV-cache compression via retrieval heads","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"desc":{"value":"Efficient KV-cache compression through retrieval heads.","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"claims":{"value":"Training-free KV cache compression that preserves all token information using head-specialized strategies — full caches for retrieval heads, discarding remote tokens in non-retrieval heads, with compensation tokens; over 70% reduction in KV cache size; 3X reduction under general cases; no noticeable performance impact; FlashAttention-compatible","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"modalities":{"value":"text","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"created":{"value":"2024-07","citation":"https://arxiv.org/abs/2407.15891","status":"real-data","tier":"T2","last_verified_at":"2026-05-12","source_class":"paper"},"latest-release":{"value":"not applicable — not OSS","citation":null,"status":"not-applicable","tier":"T3","last_verified_at":"2026-05-06"},"license":{"value":"not applicable — not OSS","citation":null,"status":"not-applicable","tier":"T3"},"gh":{"value":"not applicable — no GitHub repo","citation":null,"status":"not-applicable","tier":"T3","last_verified_at":"2026-05-06"},"mindshare":{"value":"searched not found","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"depth-floor-reached","tier":"T2","last_verified_at":"2026-05-12","source_class":"vendor"},"citations":{"value":"searched not found","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"depth-floor-reached","tier":"T2","last_verified_at":"2026-05-12","source_class":"vendor"},"funding":{"value":"not applicable — not commercial","citation":null,"status":"not-applicable","tier":"T3","last_verified_at":"2026-05-06"},"customers":{"value":"not applicable — not commercial","citation":null,"status":"not-applicable","tier":"T3"},"pricing":{"value":"not applicable — not commercial","citation":null,"status":"not-applicable","tier":"T3"},"compliance":{"value":"not applicable — not commercial","citation":null,"status":"not-applicable","tier":"T3"},"data-handling":{"value":"not applicable — not commercial","citation":null,"status":"not-applicable","tier":"T3"},"deployment":{"value":"not applicable — not a deployable product","citation":null,"status":"not-applicable","tier":"T3"},"hq":{"value":"not applicable — not a company","citation":null,"status":"not-applicable","tier":"T3"},"founders":{"value":"Hanlin Tang, Yang Lin, Jing Lin, Qingsen Han et al. (arXiv 2407.15891)","citation":"https://arxiv.org/abs/2407.15891","status":"real-data","tier":"T2","source_class":"paper"},"volume":{"value":"no scale claim — research","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"perf":{"value":">70% reduction in KV-cache size with negligible performance impact, training-free.","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"repro":{"value":"Too recent","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"code-release":{"value":"Paper only","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"api-surface":{"value":"not applicable — research paper","citation":null,"status":"not-applicable","tier":"T3"},"latency":{"value":"not applicable — research paper","citation":null,"status":"not-applicable","tier":"T3"},"throughput":{"value":"not applicable — research paper","citation":null,"status":"not-applicable","tier":"T3"},"backend-storage":{"value":"not applicable — research paper","citation":null,"status":"not-applicable","tier":"T3"},"multi-tenancy":{"value":"not applicable — research paper","citation":null,"status":"not-applicable","tier":"T3"},"encryption":{"value":"not applicable — research paper","citation":null,"status":"not-applicable","tier":"T3"},"sso-rbac":{"value":"not applicable — research paper","citation":null,"status":"not-applicable","tier":"T3"},"embedding-model":{"value":"not applicable — research paper","citation":null,"status":"not-applicable","tier":"T3"},"consistency":{"value":"strong","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"versioning":{"value":"none","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"tombstoning":{"value":"hard-delete supported","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"schema-evolution":{"value":"schema locked","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"namespace":{"value":"not applicable","citation":null,"status":"not-applicable","tier":"T3"},"contradiction":{"value":"not applicable","citation":null,"status":"not-applicable","tier":"T3"},"forgetting":{"value":"LRU eviction","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"mcp-support":{"value":"not applicable — research paper, no deployed product","citation":null,"status":"not-applicable","tier":"T3"},"a2a-support":{"value":"not applicable — research paper, no deployed product","citation":null,"status":"not-applicable","tier":"T3"},"otel":{"value":"not applicable — research paper, no deployed product","citation":null,"status":"not-applicable","tier":"T3"},"webhooks":{"value":"not applicable — research paper, no deployed product","citation":null,"status":"not-applicable","tier":"T3"},"import-export":{"value":"not applicable — research paper, no deployed product","citation":null,"status":"not-applicable","tier":"T3"},"integration-count":{"value":"not applicable — wrong section","citation":null,"status":"not-applicable","tier":"T3"},"orchestration":{"value":"not applicable — wrong section","citation":null,"status":"not-applicable","tier":"T3"},"programmatic-control":{"value":"not applicable — wrong section","citation":null,"status":"not-applicable","tier":"T3"},"vendor-benchmarks":{"value":"not applicable — wrong section","citation":null,"status":"not-applicable","tier":"T3","last_verified_at":"2026-05-06"},"pricing-specifics":{"value":"not applicable — wrong section","citation":null,"status":"not-applicable","tier":"T3"},"agent-abstraction":{"value":"not applicable — wrong section","citation":null,"status":"not-applicable","tier":"T3"},"memory-primitives":{"value":"not applicable — wrong section","citation":null,"status":"not-applicable","tier":"T3"},"llm-lock":{"value":"not applicable — wrong section","citation":null,"status":"not-applicable","tier":"T3"},"runtimes":{"value":"not applicable — wrong section","citation":null,"status":"not-applicable","tier":"T3"},"session-handling":{"value":"not applicable — wrong section","citation":null,"status":"not-applicable","tier":"T3"},"validated-verticals":{"value":"not applicable - research paper","citation":null,"status":"not-applicable","tier":"T3"},"time-to-running":{"value":"not applicable - research paper","citation":null,"status":"not-applicable","tier":"T3"},"anti-fit":{"value":"not applicable - research paper","citation":null,"status":"not-applicable","tier":"T3"},"optimised-for":{"value":"not applicable - research paper","citation":null,"status":"not-applicable","tier":"T3"},"adjacent-infrastructure":{"value":"not applicable - research paper","citation":null,"status":"not-applicable","tier":"T3"},"pros":{"value":"Aggressive KV pruning for long context with retention of high-importance tokens.","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"cons":{"value":"Research-stage; pruning quality bounded by importance heuristic.","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"links":{"value":"OpenReview","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"real-data","tier":"T2","source_class":"vendor"},"obs-langsmith":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"obs-opentelemetry":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"obs-datadog":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"obs-helicone":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"obs-weave":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"obs-langfuse":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"obs-arize":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"obs-custom":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"cost-token-budget":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"cost-prompt-caching":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"cost-semantic-caching":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"cost-batching":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"cost-model-routing":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"cost-streaming-only":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"cost-observability-cost-attribution":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"eval-langsmith-evals":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"eval-braintrust":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"eval-weights-and-biases-agent":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"eval-helicone-evals":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"eval-custom-test-harness":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"eval-human-loop":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"eval-production-traffic-replay":{"value":"","citation":null,"status":"no-data","tier":"T3","last_verified_at":"2026-05-14"},"commit-trajectory":{"value":"not applicable — no GitHub repo","citation":null,"status":"not-applicable","tier":"T3","last_verified_at":"2026-05-14"},"citation-trajectory":{"value":"not applicable — non-paper row","citation":null,"status":"not-applicable","tier":"T3","last_verified_at":"2026-05-14"},"download-trajectory":{"value":"not applicable — not published as a package","citation":null,"status":"not-applicable","tier":"T3","last_verified_at":"2026-05-14"},"cost-input-usd-per-mtok":{"value":"not applicable — not commercial","citation":null,"status":"not-applicable","tier":"T3"},"cost-output-usd-per-mtok":{"value":"not applicable — not commercial","citation":null,"status":"not-applicable","tier":"T3"},"cost-tier":{"value":"not applicable — not commercial","citation":null,"status":"not-applicable","tier":"T3"},"cost-pricing-model":{"value":"not applicable — not commercial","citation":null,"status":"not-applicable","tier":"T3"},"cost-last-verified":{"value":"not applicable — not commercial","citation":null,"status":"not-applicable","tier":"T3"},"capability-composite-score":{"value":"not applicable — research paper / KV-cache method, not a runnable capable system","citation":null,"status":"not-applicable","tier":"T3"},"capability-band":{"value":"not applicable — research paper / KV-cache method, not a runnable capable system","citation":null,"status":"not-applicable","tier":"T3"},"capability-benchmark-sources":{"value":"not applicable — research paper / KV-cache method, not a runnable capable system","citation":null,"status":"not-applicable","tier":"T3"},"capability-last-verified":{"value":"not applicable — research paper / KV-cache method, not a runnable capable system","citation":null,"status":"not-applicable","tier":"T3"},"use-case-tags":{"value":"not applicable — research method paper, no deployable system","citation":null,"status":"not-applicable","tier":"T3"},"use-case-anti-tags":{"value":"not applicable — research method paper, no deployable system","citation":null,"status":"not-applicable","tier":"T3"},"adoption-accelerator":{"value":"","citation":null,"status":"no-data","tier":"T3"},"architecture-deployment-locus":{"value":"","citation":null,"status":"no-data","tier":"T3"},"licensing-bucket":{"value":"not applicable — research paper","citation":"https://openreview.net/forum?id=tkiZQlL04w","status":"not-applicable","tier":"T2","source_class":"vendor"},"positioning-platform-vs-product":{"value":"","citation":null,"status":"no-data","tier":"T3"},"taxonomy-three-layer-rollup":{"value":"agent-infra","citation":"https://github.com/MrPeppersDev/agent-infrastructure-landscape/blob/main/docs/schema/section-to-three-layer-rollup.json","status":"real-data","tier":"T1","source_class":"repo"},"developer-activity-opendigger":{"value":"","citation":null,"status":"no-data","tier":"T3"},"freshness-date":{"value":"2026-10-23","citation":"https://mrpeppersdev.github.io/agent-infrastructure-landscape/systems/razorattention--openreview-net","status":"real-data","tier":"T2","source_class":"catalog"},"compliance-certifications":{"value":"","citation":null,"status":"no-data","tier":"T3"},"sdk-languages":{"value":"","citation":null,"status":"no-data","tier":"T3"},"protocol-support":{"value":"","citation":null,"status":"no-data","tier":"T3"},"maturity":{"value":"","citation":null,"status":"no-data","tier":"T3"}},"corroboration":{"cited_cells":19,"third_party_cells":0},"edges":{"outgoing":[],"incoming":[]},"history":[],"license":"CC-BY-4.0","source":"https://github.com/MrPeppersDev/agent-infrastructure-landscape/blob/main/data/landscape.json"}