内容库 / 论文

arXiv:2608.22752 · Agent 智能体

The Compaction Cliff in Long-Running AI Agent Memory

长时运行 AI Agent 记忆中的压缩悬崖

The Compaction Cliff in Long-Running AI Agent Memory

类型：arxiv
标识：2608.22752
链接：http://arxiv.org/abs/2608.22752v1
主分类：agent
形态：method
TLDR：A safety rule and an episodic log compete for the same tokens in an AI agent's context. When the budget overflows, both are summarized at the same rate; only the rule needs exact wording to remain enforceable. On 20 production agent configurations, Claude Code's /compact prompt on Sonnet 4.6 preserves 53\% of safety rules after one compaction round and 10\% after five. We name this the Compaction Cliff. We address it with Knowledge Triage, a framework that classifies each line of an agent's knowledge base by type and routes each type through its own retention policy. Three deterministic operat
副分类：risk
待LLM分类：否
标题中文：长时运行 AI Agent 记忆中的压缩悬崖
TLDR中文：一条安全规则与一段情景日志在同一 AI Agent 上下文中争夺 token。当预算溢出时，二者以相同速率被压缩；但只有规则需要精确措辞才能保持可执行性。在 20 种生产环境 Agent 配置下，Claude Code 基于 Sonnet 4.6 的 /compact 提示在一轮压缩后保留 53% 的安全规则，五轮后仅保留 10%。我们将此现象命名为"压缩悬崖"（Compaction Cliff）。我们提出 Knowledge Triage 框架，通过对 Agent 知识库的每一行按类型分类，并为每类配置独立的保留策略来解决该问题。三种确定性 operat
来源文件：
/inbox/tom/_candidates/2026-08-25-agent-rag-longcontext-candidates.json

相关 · 按语义匹配

笔记 Tom 文献雷达 · AI Agent / RAG / 长上下文 · 2026-07-30（午后版）

笔记 flyP 轻量精读 · Beyond Memory: A Transactional Continuity Kernel for Long-Lived AI Agents（arXiv:2608.11632 · cs.MA · 8-12 v1）

笔记 Tom 文献雷达 · AI Agent × RAG × Long Context · 2026-06-24

笔记 2026-07-09 研究笔记：RAG 优化与推理引擎实战 + Substack AI Agents Stack（**重写版 v2 · 2026-07-09 21:10 CST**）

笔记 flyP 反方审稿 · Beyond Memory · 8-13 15:50 7 条基础反方 → 8-15 收敛为 3 条硬反方

笔记 SpecBench: Measuring Reward Hacking in Long-Horizon Coding Agents — 短精读

arXiv

2608.22752

TLDR

一条安全规则与一段情景日志在同一 AI Agent 上下文中争夺 token。当预算溢出时，二者以相同速率被压缩；但只有规则需要精确措辞才能保持可执行性。在 20 种生产环境 Agent 配置下，Claude Code 基于 Sonnet 4.6 的 /compact 提示在一轮压缩后保留 53% 的安全规则，五轮后仅保留 10%。我们将此现象命名为"压缩悬崖"（Compaction Cliff）。我们提出 Knowledge Triage 框架，通过对 Agent 知识库的每一行按类型分类，并为每类配置独立的保留策略来解决该问题。三种确定性 operatA safety rule and an episodic log compete for the same tokens in an AI agent's context. When the budget overflows, both are summarized at the same rate; only the rule needs exact wording to remain enforceable. On 20 production agent configurations, Claude Code's /compact prompt on Sonnet 4.6 preserves 53\% of safety rules after one compaction round and 10\% after five. We name this the Compaction Cliff. We address it with Knowledge Triage, a framework that classifies each line of an agent's knowledge base by type and routes each type through its own retention policy. Three deterministic operat

相关解读

arXiv ↗ 中文全文