返回
原文
返回
AI 评分 47/100在 X 查看原推
导出 Markdown

KVMEM：通过分页旧 KV 状态让百万 token 级智能体记忆变得实用

Rohan Paul · @rohanpaul_ai · X·2026-09-23 09:36·14分钟前
AI 导读

KVMEM 通过分页旧 KV 状态，让百万 token 级智能体记忆变得实用。在 DeepSWE 上用 Qwen3.8-27B 将 Pass@1 从 43.8% 提升至 48.4%，优于仅用压缩方案，恢复速度在受控基准上比 Compact+RAG 快 11.4–53.8×。

Rohan Paul@rohanpaul_ai
47AI 编辑部评分，满分 100

KVMEM：通过分页旧 KV 状态让百万 token 级智能体记忆变得实用

2026-09-23 09:36· 14分钟前
AI 导读

KVMEM 通过分页旧 KV 状态，让百万 token 级智能体记忆变得实用。在 DeepSWE 上用 Qwen3.8-27B 将 Pass@1 从 43.8% 提升至 48.4%，优于仅用压缩方案，恢复速度在受控基准上比 Compact+RAG 快 11.4–53.8×。

A model’s context window does not have to be an agent’s workspace limit.

KVMEM makes million-token agent memory practical by paging old KV state, cutting recall cost while beating compaction on task success.

Long-running agents eventually fill their context. Today, they usually compress old history into summaries or fetch old text again. Summaries can forget details. Fetching text makes the model process the same history again.

KVMEM keeps the work the model already did. Old context stays as reusable KV cache in GPU memory, RAM, or NVMe, and the system pulls back only the pieces needed for the current step.

On DeepSWE with Qwen3.8-27B, Pass@1 rose from 43.8% to 48.4% versus compaction-only. Across controlled benchmarks, recovery was 11.4–53.8× faster than Compact+RAG.

On a laptop with a 24 GB RTX 5090 GPU, KVMEM supported a 1M-token workspace at about 50 tokens/s.

This is not a 1M-token active prompt. Each step still sees only a bounded slice.

来源：Rohan Paul· x.com

导出 Markdown