Back to Arena

MemoryScope

by Alibaba ModelScope

System Card

OrganizationAlibaba ModelScope
Released2024-09
Architecturehierarchical-summary / Tiered consolidation with retrieval
DetailsTiered store of raw episodes, mid-term summaries, and long-term semantic notes. Background workers consolidate older tiers; retrieval queries the tier best matched to the question.
Parameters—
Domainagent-memoryepisodic-sessionpersonalization
Open SourceYes
CodeRepository
tieredconsolidationbackground-worker

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
LoCoMo
66.620p
MemoryBank
74.339p
Cross-Session Memory
1/1
LongMemEval
63.718p
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
73.263p
HotpotQA
71.249p
Agent Task Memory
1/1
AgentBench-Mem
68.919p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:MemoryScope evalsMemoryScope (modelscope/MemoryScope); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)MemoryScope (modelscope/MemoryScope); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)MemoryScope (modelscope/MemoryScope); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)MemoryScope (modelscope/MemoryScope); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)