Back to Arena

ReMe

by ModelScope (Alibaba)

System Card

OrganizationModelScope (Alibaba)
Released2025-01
Architecturehybrid / File-based + vector memory with ReAct compactor
DetailsDual-mode memory: ReMeLight writes Markdown/JSONL files (MEMORY.md, daily journals, dialog logs, tool-result cache), while vector store uses hybrid BM25 + embeddings. A ReAct compactor summarizes into Goal/Progress/Decisions/Next-Steps.
Parameters—
Domainagent-memorylong-context
Open SourceYes
CodeRepository
file-memoryreact-compactorbm25-hybridalibabalocomo

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Long-Context Retrieval
2/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
603p
∞Bench
7522p
Multi-Turn Recall
1/2
LoCoMo
7335p
MemoryBank
no data
Cross-Session Memory
0/1
LongMemEval
no data
Multi-Hop QA
2/3
BABILong
79.886p
MultiHop-RAG
no data
HotpotQA
76.278p
Agent Task Memory
1/1
AgentBench-Mem
7226p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:ReMe (modelscope/ReMe); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)ReMe (modelscope/ReMe); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)ReMe (modelscope/ReMe); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)ReMe (modelscope/ReMe); evaluated on InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens (Tsinghua / OpenBMB, 2402)ReMe (modelscope/ReMe); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)ReMe (modelscope/ReMe); evaluated on LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (Tsinghua KEG, 2308)