Figure: the 2026 KV-cache landscape, mapped by the decision each system optimises, in four bands plus an industry timeline. The field did not leave a gap; it specialised. The earlier framing on this figure — that KV-cache-over-CXL was uniformly unsophisticated and that nobody had built attention-aware retention — is retired outright, not softened. Commercial hardware available today: Samsung CMM-D up to 512 GB per module and CMM-B up to 16 TB per box; the XConn Apollo 256-lane CXL switch with MemVerge GISMO pool demos; the Astera Labs Leo smart memory controller; and Marvell Structera near-memory compute. Software and research systems each optimise a different sub-decision: vLLM does GPU KV allocation and paging via PagedAttention, the baseline every offload layer sits on; LMCache does KV persistence, reuse and movement across tiers and engines, with pinning, lookup, compression, batched movement, compute and I/O pipelining, cross-engine transfer and prefill-decode disaggregation; TraCT does rack-scale KV sharing and transfer over CXL, implemented on NVIDIA Dynamo with direct GPU-to-CXL operations behind a rack-wide prefix-aware cache — it is already a CXL-native KV-cache system, not a gap waiting to be filled; PNM-KV selects which KV pages are needed before paying the movement cost, using a processing-near-memory accelerator in or near CXL memory, and reports up to 21.9 times throughput improvement in its own benchmarks, cited as their reported result and not independently verified; Tutti removes the CPU from the KV I/O critical path for GPU-to-NVMe object movement, reporting 78.3 percent time-to-first-token reduction against GDS-enabled SSD-backed LMCache under strict service objectives; CacheWise predicts future KV reuse from agent and tool-call metadata, reporting 2 to 2.6 times fewer evictions and up to 3.5 times faster agent session completion. Alongside them, the model-runtime line of work this package builds on rather than competes with: H2O, SnapKV, Scissorhands, PyramidKV, Quest, Ada-KV, HeadKV, InfiniGen, LayerKV, KVPR and PagedAttention. This package's own position is an open design question, not a shipped system: whether relevance estimation, reuse prediction, residency and movement can be decided jointly by one KV-State Control Plane — SELECT, PLACE, MOVE, ATTEND, OBSERVE — instead of separately by each layer above. Only one piece of that has any empirical result, and it is simulated and provisional: inside PLACE, letting identity-keyed policy state survive HBM eviction is worth plus 34.15 percentage points to LFU, from 15.56 to 49.71 percent, and plus 24.85 to EMA alpha 0.01, from 31.54 to 56.39 percent, at a 32 gibibyte budget, while EMA alpha 0.15 is unchanged at minus 0.21 percentage points. Retracted from this figure and not replaced: the modelled 97 percent hit rate, the 16 times user capacity, alpha = 0.2, the per-head eviction queues, and the RoPE prefetch window — RoPE is a rotation applied to Q and K at compute time and carries no information about which blocks are needed next. Every head of every layer is required for every token, so evicting a head produces wrong outputs, not a cache miss; H2O and SnapKV evict head-token pairs, and the token axis is not optional. No CXL hardware was measured. Simulation figures are provisional, because a single data-structure defect in Revision 1 inverted four of five findings. The timeline runs 2022 to 2026, colour-coded green for hardware, amber for software, purple for research and blue for standards.

The 2026 KV-Cache Landscape

The field did not leave a gap — it specialised. Mapped by the decision each system optimises.

Evidence labels. Third-party capabilities are vendor specifications or the systems' own published claims, cited not reproduced — PNM-KV's 21.9×, Tutti's 78.3% TTFT reduction and CacheWise's 2–2.6× fewer evictions are their reported results, not independently verified here. The persistence result is Simulated (provisional): a single data-structure defect in Revision 1 inverted four of five findings, so Revision 2 is provisional, not settled. Capacity arithmetic is an Analytical model (320 KiB/token, 2.50 GiB at 8K, 40.0 GiB at 128K, crossover ≈ 427,000 tokens). LRFU attribution is External literature. Retracted from this figure: the 97% hit rate, 16× user capacity, 36% CapEx / $160,000 → $103,000, α = 0.2, per-head eviction, RoPE-aware prefetch, and every “nobody has” framing. No CXL hardware was measured. Canonical numbers v6.0.