KVMem pages KV state to give agents million-token workspaces on a consumer GPU
rohanpaul_ai · x · 2026-09-23
KVMem is a KV-context virtualization system that keeps overflowed agent workspace history as paged KV state across GPU memory, host RAM, and NVMe, instead of compacting it into summaries or re-fetching text.
- On DeepSWE with Qwen3.8-27B, Pass@1 rises from 43.8% (compaction-only) to 48.4%
- History recovery is 11.4–53.8× faster than Compact+RAG across benchmarks
- On a laptop with a 24GB RTX 5090, it sustains a 1M-token workspace at 50 tokens/s; each step still sees only a bounded slice
Related event: KVMem Virtualizes Million-Token Agent Workspaces on Consumer GPUs(2 posts)→
More from coding & agent
- Stanford-led Terminal-Bench-Science debuts: top model solves just 30% of 70 expert tasks — geoffwolfe · 2026-09-23
- Moda's agent observability weekly: Jev for sharper signals, whole-conversation analysis for long-running agents — KlausCodes · 2026-09-23
- Opus 5.5 builds a Lanterns Festival scene in one prompt, using just 13% of a weekly Claude Max budget — Silver-Chipmunk7744 · 2026-09-23
- stuntd: a local proxy that learns your LLM decisions and serves them at 22ms without an API key — Inevitable-Log5414 · 2026-09-23
- Metriqual: an infra layer that lets AI agents survive model outages by persisting their state — its_vayishu · 2026-09-23
- Adding OAuth to a Sonos MCP server: 6-digit codes beat redirects, discovery metadata matters most — Mean-Gazelle5347 · 2026-09-23