Generative Agents

by Stanford / Google

System Card

OrganizationStanford / Google

Released2023-04

Architecturehierarchical-summary / Memory stream + reflection tree + planning

DetailsExtends an LLM with a complete natural-language record of agent experiences (memory stream), synthesizes them over time into hierarchical reflections (non-leaf nodes of a reflection tree), and retrieves them dynamically by recency, importance, and relevance to plan behavior.

Parameters—

Domainagent-memoryepisodic-sessionlifelong-learning

Open SourceYes

PaperView Paper

uist-2023reflectionsmallvillememory-streamsimulation

Capability Profile

Benchmark Scores

6 of 14 benchmarks

Long-Context Retrieval

0/5

RULER

no data

NIAH

no data

LooGLE

no data

LongBench

no data

∞Bench

no data

Multi-Turn Recall

2/2

7232p

65.416p

Cross-Session Memory

1/1

71.126p

Multi-Hop QA

2/3

71.514p

MultiHop-RAG

no data

7254p

Agent Task Memory

1/1

7226p

Personalization

0/1

PerLTQA

no data

Factuality / Grounding

0/1

RAGAS

no data

Sources:Generative Agents paper (arXiv:2304.03442); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)Generative Agents paper (arXiv:2304.03442); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)Generative Agents paper (arXiv:2304.03442); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)Generative Agents paper (arXiv:2304.03442); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)Generative Agents paper (arXiv:2304.03442); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)Generative Agents paper (arXiv:2304.03442); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)