Back to Arena

MemoRAG

by BAAI / Qhjqhj00

System Card

OrganizationBAAI / Qhjqhj00
Released2024-09
Architectureexternal-memory-network / Long-context "memory model" + retriever + generator
DetailsA 1M-token memory LM (Qwen2/Mistral fine-tunes) builds global understanding of a corpus, generates query-specific clues, which guide a bge-m3 retriever. The generator (memory LM itself or external) produces final answers using memory-inspired knowledge discovery.
Parameters—
Domainrag-retrievallong-context
Open SourceYes
PaperView Paper
CodeRepository
memory-model1m-tokensbge-m3theweb-conf-2025

Capability Profile

Benchmark Scores

3 of 14 benchmarks
Long-Context Retrieval
2/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
44.42p
∞Bench
24.50p
Multi-Turn Recall
0/2
LoCoMo
no data
MemoryBank
no data
Cross-Session Memory
0/1
LongMemEval
no data
Multi-Hop QA
1/3
BABILong
no data
MultiHop-RAG
no data
HotpotQA
54.86p
Agent Task Memory
0/1
AgentBench-Mem
no data
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:arXiv:2409.05591 Table 1 — Mistral-7B-v0.2-32K memory + Phi-3-mini-128K generator; avg of NarrativeQA 27.5, Qasper 43.9, MultiFieldQA 52.2, MuSiQue 33.9, 2WikiMQA 54.1, HotpotQA 54.8arXiv:2409.05591 Table 1 — MemoRAG on HotpotQA (via LongBench)arXiv:2409.05591 Table 1 — avg of MultiNews 26.3, GovReport 32.9, En.SUM 15.7, En.QA 22.9