Back to Arena

R3Mem

by HKUST (2025)

System Card

OrganizationHKUST (2025)
Released2025-02
Architecturekv-cache-extension / Reversible hierarchical context compression
DetailsVirtual memory tokens compress text hierarchically at document/paragraph/entity levels. A reversible architecture (adapter-tuned on a frozen Transformer) reconstructs raw input by inverting the compression for retrieval.
Parameters—
Domainlong-contextagent-memory
Open SourceNo
PaperView Paper
acl-2025reversiblevirtual-tokenshierarchical

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
81.975p
Multi-Turn Recall
1/2
LoCoMo
62.314p
MemoryBank
no data
Cross-Session Memory
0/1
LongMemEval
no data
Multi-Hop QA
2/3
BABILong
75.857p
MultiHop-RAG
no data
HotpotQA
60.313p
Agent Task Memory
1/1
AgentBench-Mem
59.23p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:R3Mem paper (arXiv:2502.15957); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)R3Mem paper (arXiv:2502.15957); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)R3Mem paper (arXiv:2502.15957); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)R3Mem paper (arXiv:2502.15957); evaluated on InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens (Tsinghua / OpenBMB, 2402)R3Mem paper (arXiv:2502.15957); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)R3Mem paper (arXiv:2502.15957); evaluated on LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (Tsinghua KEG, 2308)