{"id":"h2o","name":"H2O","url":null,"page":"https://mrpeppersdev.github.io/agent-infrastructure-landscape/systems/h2o","family":"memory-research","section":"Recent method papers — theorized, no distinct product","sections":["Recent method papers — theorized, no distinct product"],"tier":3,"last_verified_at":"2026-10-28","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":"deterministic","primary":true,"reason":null}],"conflict":[{"value":"n/a","primary":true,"reason":"not applicable (KV eviction; no user write)"}]},"cells":{"type":{"value":"Heavy-hitter oracle KV eviction","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"desc":{"value":"Heavy-hitter oracle for efficient generative inference. NeurIPS 2023.","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"claims":{"value":"Heavy-Hitter Oracle for KV-cache eviction balancing recent and high-impact tokens. Headline: up to 29x throughput on OPT-6.7B/OPT-30B with 20% heavy hitters retained, 1.9x latency reduction; baselines: DeepSpeed Zero-Inference, Hugging Face Accelerate, FlexGen; primary datasets: OPT/LLaMA/GPT-NeoX evaluated on HELM and lm-eval-harness (COPA, MathQA, OpenBookQA, PiQA, RTE, Winogrande, XSUM, CNN/Daily Mail), AlpacaEval, MT-Bench.","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"modalities":{"value":"text","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"created":{"value":"2023-06-24 (arxiv 2306.14048 submitted June 24 2023; NeurIPS 2023)","citation":"https://arxiv.org/abs/2306.14048","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":"654 cites · 19.0/mo · 516* / 80 forks · 92 influential SS cites","citation":"https://api.semanticscholar.org/graph/v1/paper/arXiv:2306.14048","status":"real-data","tier":"T2","last_verified_at":"2026-05-12","source_class":"paper"},"citations":{"value":"searched not found","citation":"https://www.semanticscholar.org/paper/H2O:-Heavy-Hitter-Oracle-for-Efficient-Generative-Zhang-Sheng/e586a4591ba0303b769f2c07cbddaf1899cb72e4","status":"depth-floor-reached","tier":"T2","last_verified_at":"2026-05-12","source_class":"paper"},"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":"no data — no URL in landscape HTML; searched NeurIPS 2023 H2O heavy-hitter oracle","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"volume":{"value":"no scale claim — research (NeurIPS 2023)","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"perf":{"value":"29x OPT-6.7B/30B throughput (20% heavy hitters) 1.9x latency reduction","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"repro":{"value":"Independently reproduced: integrated into vLLM and HuggingFace transformers; widely cited as KV-cache eviction baseline","citation":"https://github.com/FMInference/H2O","status":"real-data","tier":"T1","source_class":"repo"},"code-release":{"value":"Code public + weights closed — https://github.com/FMInference/H2O · stale (last commit 2024-08) · 516* / 80 forks · no license","citation":"https://github.com/FMInference/H2O","status":"real-data","tier":"T1","source_class":"repo"},"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://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"versioning":{"value":"none","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"tombstoning":{"value":"hard-delete supported","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"schema-evolution":{"value":"schema locked","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"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://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"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":"Heavy-hitter oracle KV eviction — keeps tokens with high attention scores, evicts the rest; NeurIPS 2023.","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"cons":{"value":"Heavy-hitter heuristic can miss tokens that become important later; eviction is irreversible.","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"links":{"value":"NeurIPS 2023","citation":"https://arxiv.org/abs/2306.14048","status":"real-data","tier":"T2","source_class":"paper"},"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":"[{\"year\":2024,\"cum\":7,\"infl\":1},{\"year\":2025,\"cum\":11,\"infl\":1}]","citation":"https://www.semanticscholar.org/paper/e586a4591ba0303b769f2c07cbddaf1899cb72e4","status":"real-data","tier":"T2","last_verified_at":"2026-05-14","source_class":"paper"},"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 on KV-cache eviction policy, not a capable system","citation":null,"status":"not-applicable","tier":"T3"},"capability-band":{"value":"not applicable — research paper on KV-cache eviction policy, not a capable system","citation":null,"status":"not-applicable","tier":"T3"},"capability-benchmark-sources":{"value":"not applicable — research paper on KV-cache eviction policy, not a capable system","citation":null,"status":"not-applicable","tier":"T3"},"capability-last-verified":{"value":"not applicable — research paper on KV-cache eviction policy, not a capable system","citation":null,"status":"not-applicable","tier":"T3"},"use-case-tags":{"value":"not applicable — research paper on KV eviction, no deployable system","citation":null,"status":"not-applicable","tier":"T3"},"use-case-anti-tags":{"value":"not applicable — research paper on KV eviction, 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":null,"status":"not-applicable","tier":"T3"},"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":"2026-10-28","citation":"https://mrpeppersdev.github.io/agent-infrastructure-landscape/systems/h2o","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":21,"third_party_cells":0},"edges":{"outgoing":[],"incoming":[{"source":"namms--arxiv-2410-13166","target":"h2o","type":"cites","evidence":"S2 isInfluential citation","citation":"https://api.semanticscholar.org/graph/v1/paper/arXiv:2410.13166","isInfluential":true}]},"history":[],"license":"CC-BY-4.0","source":"https://github.com/MrPeppersDev/agent-infrastructure-landscape/blob/main/data/landscape.json"}