Back to Arena

RMM

by Google / UCSB (2025)

System Card

OrganizationGoogle / UCSB (2025)
Released2025-03
Architectureagentic-workflow / Prospective + retrospective reflection with online RL retrieval
DetailsCombines Prospective Reflection (dynamic multi-granularity summarization into a personalized memory bank) with Retrospective Reflection, which refines retrieval online via reinforcement learning from LLM-cited evidence.
Parameters—
Domainpersonalizationagent-memoryepisodic-session
Open SourceNo
PaperView Paper
acl-2025rlreflectionmulti-granularlongmemeval

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
72.735p
MemoryBank
74.741p
Cross-Session Memory
1/1
LongMemEval
70.423p
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
7579p
HotpotQA
73.964p
Agent Task Memory
1/1
AgentBench-Mem
7226p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:arXiv:2503.08026 Table 1 (ACL 2025) — RMM with GTE retriever; baseline GTE RAG 63.6%. >10% improvement over no-memory baselineRMM paper (arXiv:2503.08026); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)RMM paper (arXiv:2503.08026); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)RMM paper (arXiv:2503.08026); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)RMM paper (arXiv:2503.08026); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)RMM paper (arXiv:2503.08026); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)