Robots Atlas>ROBOTS ATLAS

AI Agent Architecture — ReAct, Memory, Planning and Multi-Agent Systems · Agent Memory — From Context Window to Vector Store

Memory stream and retrieval scoring: recency × importance × relevance

Agent Memory — From Context Window to Vector Store

Introduction

Generative Agents (Park et al. 2023) defined the memory stream architecture: each event is stored with a timestamp and importance score, and retrieval is driven by a weighted product of three signals: recency (freshness), importance, and relevance (semantic match). This lesson dissects each component, the combining formula, how importance is computed (LLM-based scoring), and the practical design implications.