Back to Arena

MemR3

by 2025 (December submission)

System Card

Organization2025 (December submission)
Released2025-12
Architectureagentic-workflow / LangGraph closed-loop retrieve-reflect-answer router
DetailsAutonomous memory-retrieval controller built on LangGraph. Router chooses among retrieve/reflect/answer actions; a global evidence-gap tracker monitors what evidence is still missing. Agnostic to backend retrievers (vector, graph, hybrid).
Parameters—
Domainagent-memoryrag-retrieval
Open SourceNo
PaperView Paper
langgraphclosed-loopevidence-gaplocomo

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Long-Context Retrieval
1/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
603p
∞Bench
no data
Multi-Turn Recall
1/2
LoCoMo
86.898p
MemoryBank
no data
Cross-Session Memory
1/1
LongMemEval
83.994p
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
7579p
HotpotQA
8298p
Agent Task Memory
1/1
AgentBench-Mem
7226p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:arXiv:2512.20237 Table 1 — GPT-4.1-mini + RAG backbone, LLM-as-Judge overallMemR3 paper (arXiv:2512.20237); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)MemR3 paper (arXiv:2512.20237); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)MemR3 paper (arXiv:2512.20237); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)MemR3 paper (arXiv:2512.20237); evaluated on LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (Tsinghua KEG, 2308)MemR3 paper (arXiv:2512.20237); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)