Back to Arena

Memformer

by UC Santa Barbara / Amazon (Wu, Lan, Liu, et al.)

System Card

OrganizationUC Santa Barbara / Amazon (Wu, Lan, Liu, et al.)
Released2020-10
Architectureexternal-memory-network / External dynamic memory + memory replay BPTT
DetailsUses an external dynamic memory module to encode/retrieve past information in linear-time constant-space. Introduces Memory Replay Back-Propagation (MRBP) for efficient long-range BPTT.
Parameters—
Domainlong-contextagent-memory
Open SourceYes
PaperView Paper
CodeRepository
aacl-findings-2022external-memorydynamiclinear-time

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
78.750p
Multi-Turn Recall
1/2
LoCoMo
65.717p
MemoryBank
no data
Cross-Session Memory
0/1
LongMemEval
no data
Multi-Hop QA
2/3
BABILong
74.948p
MultiHop-RAG
no data
HotpotQA
6839p
Agent Task Memory
1/1
AgentBench-Mem
65.815p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:Memformer paper (arXiv:2010.06891); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)Memformer paper (arXiv:2010.06891); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)Memformer paper (arXiv:2010.06891); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)Memformer paper (arXiv:2010.06891); evaluated on InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens (Tsinghua / OpenBMB, 2402)Memformer paper (arXiv:2010.06891); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)Memformer paper (arXiv:2010.06891); evaluated on LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (Tsinghua KEG, 2308)