Back to Arena

Agent Workflow Memory

by CMU (Wang, Mao, Fried, Neubig)

System Card

OrganizationCMU (Wang, Mao, Fried, Neubig)
Released2024-09
Architectureagentic-workflow / Induced-workflow (routine) memory
DetailsInduces commonly reused routines (workflows) from training examples or on-the-fly from test queries, then selectively injects them to guide subsequent generations. Workflows are context-abstracted sub-routines.
Parameters—
Domainagent-memorylifelong-learning
Open SourceYes
PaperView Paper
CodeRepository
web-agentworkflowsroutinesmind2webwebarena

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Data Transparency:6 estimated
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
LoCoMo
8191pEstimated
MemoryBank
77.965pEstimated
Cross-Session Memory
1/1
LongMemEval
78.768pEstimated
Multi-Hop QA
2/3
BABILong
7657pEstimated
MultiHop-RAG
no data
HotpotQA
73.663pEstimated
Agent Task Memory
1/1
AgentBench-Mem
7226pEstimated
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:Arena estimate — derived from capability profile, not independently verified