← 首页|学术|Muscle Memory: Beyond Retrieval for Personalizing Agents
cs.MA · 2608.08995 · 2026-08-10

Muscle Memory for Agents: Compile not Merely Retrieve

Pouya Ghiasnezhad Omran, Soujanya Lanka, Qin Zhang, Tanya Dixit
TL;DR:agent 记忆收敛到"存文本-推理时检索-编排器解读"的单一范式,论文主张对个性化而言编译(compile)优于检索:把重复用户意图编译成专用 specialist agent(Harvest→Analyze→Augment→Evaluate 四阶段管线),90 场景 88.9% 胜率、+2.05 个性化增益。

🎯 问题

LLM agent 记忆已收敛于单一架构模式:把经验存为文本/嵌入/反思/规则,推理时检索,让通用编排器解读。论文认为这对个性化是错误的默认——当前助手对用户强加"多轮税":反复纠正格式、深度、范围才能得到领域合适的答案。

🔬 方法

把 Muscle Memory(将重复用户意图编译为专用 specialist agents)定位为区别于检索的独立记忆范式。四阶段管线 Harvest→Analyze→Augment→Evaluate:挖掘会话历史、分离行为模式与任务模式,产出质量门控的可执行编译 specialists,配两阶段触发匹配。
章节结构(全文标题提取):
1 Introduction
2 Related Work / The Retrieval Paradigm and Its Limits
3 The Compiled Memory Position
4 A Reference Implementation
5 Empirical Evidence
6 Discussion and Limitations
7 Conclusion and Future Work / Implications and Open Problems
Acknowledgments
Declaration on Generative AI
· 2.1 Conversational Memory, Reasoning Memory, and Personalization
· 2.2 Skill Libraries and Agentic Skills
· 2.3 Multi-Agent Systems
· 2.4 LLM-as-Judge Evaluation
· 3.1 Three principles
· 3.2 Compiled vs. retrieved: where they diverge
· 3.3 When compilation is the right choice
· 4.1 Phase 1: History Harvest
· 4.2 Phase 2: Pattern Analysis and Swarm Generation
· 4.3 Phase 3: Runtime Augmentation

📊 结果

90 个 held-out 场景、5 个用户 persona:specialist 触发时 36 例赢 32(88.9%),+2.05 个性化增益、仅 -0.28 精度代价(1-4 标度)。

📝 原文摘要

▶ 原文摘要 Abstract
Memory for LLM agents has converged on a single architectural pattern: store experience as text, embeddings, reflections, or rules; retrieve at inference time; let a general-purpose orchestrator interpret what to do. This paper argues that the pattern is the wrong default for personalization. We position Muscle Memory - the practice of compiling recurring user intent into purpose-built specialist agents - as a distinct memory paradigm from retrieval, and we argue that compilation is a better fit for the workloads where current assistants impose a multi-turn tax on their users: making them repeatedly correct format, depth, and scope to obtain a domain-appropriate answer. We support the position with a reference implementation and empirical evidence. The implementation is a four-phase pipeline (Harvest $\rightarrow$ Analyze $\rightarrow$ Augment $\rightarrow$ Evaluate) that mines conversational history, separates behavioral from task patterns, and emits quality-gated executable compiled specialists with two-stage trigger matching. On 90 held-out scenarios across five user personas, the augmented assistant wins 32 of 36 cases where a specialist fires, an 88.9% win rate, with a +2.05 personalization gain and only a $-0.28$ accuracy cost on a 1-4 scale. We discuss why compilation is better suited than retrieval in this regime, what the result implies for the broader memory design space, and what open problems remain.
Deep Read · 2026-08-12高松灯 / Agent 日报
AgentMemoryPersonalizationCompilation