Voyager

by NVIDIA / Caltech / UT Austin / Stanford / ASU / UW (Wang et al.)

System Card

OrganizationNVIDIA / Caltech / UT Austin / Stanford / ASU / UW (Wang et al.)

Released2023-05

Architectureagentic-workflow / Ever-growing code skill library + curriculum

DetailsLifelong Minecraft agent with three components: automatic curriculum, ever-growing skill library of executable code, and iterative prompting with environment feedback. Skills are retrieved by embedding similarity and compounded without catastrophic forgetting.

Parameters—

Domainagent-memorylifelong-learning

Open SourceYes

PaperView Paper

WebsiteVisit

minecraftskill-libraryembodiedcurriculumcode

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

83.695p

75.753p

Cross-Session Memory

1/1

87.198p

Multi-Hop QA

2/3

78.481p

MultiHop-RAG

no data

81.598p

Agent Task Memory

1/1

7226p

Personalization

0/1

PerLTQA

no data

Factuality / Grounding

0/1

RAGAS

no data

Sources:Voyager paper (arXiv:2305.16291); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)Voyager paper (arXiv:2305.16291); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)Voyager paper (arXiv:2305.16291); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)Voyager paper (arXiv:2305.16291); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)Voyager paper (arXiv:2305.16291); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)Voyager paper (arXiv:2305.16291); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)