# RSM-full：分组记忆让受限智能体省 68% token

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-09-11 11:10
- AIHOT 分数：44
- AIHOT 链接：https://aihot.news/items/cmtwdrrvv0d9grolk213r1lgl
- 原文链接：https://x.com/rohanpaul_ai/status/2098247684620726776

## AI 摘要

RSM-full 在 AMA-Bench 约 4k prompt token 预算下，保留了完整历史输入 83% 的质量，token 成本仅为后者的 32%。它按记忆到达顺序将相关记忆分组，需要时整组召回，而非检索孤立片段。作者认为，对内存受限的智能体，先组织好过去比优化检索更有效。

## 正文

For memory-limited agents, organize the past before you optimize search: grouping related memories and keeping them together beat fancier retrieval under tight token budgets.

Most long-running agents either keep stuffing old interactions into the prompt or retrieve isolated chunks.

Both approaches get awkward once the prompt budget gets tight.

RSM-full takes a simpler route.

It groups related memories as they arrive, then brings those groups back together when the agent needs them.

On AMA-Bench at roughly 4k prompt tokens, it kept 83% of the quality of feeding the model its full history while using 32% of the token cost.