Back to Arena

Memorizing Transformer

by Google Research (Wu, Rabe, Hutchins, Szegedy)

System Card

OrganizationGoogle Research (Wu, Rabe, Hutchins, Szegedy)
Released2022-03
Architectureexternal-memory-network / Non-differentiable kNN lookup over (key,value) pairs
DetailsApproximate kNN lookup into a non-differentiable cache of recent attention (key, value) pairs. Scales the effective attention context up to 262k tokens.
Parameters—
Domainlong-contextlifelong-learning
Open SourceNo
PaperView Paper
WebsiteVisit
iclr-2022-spotlightknnnon-differentiable262k

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Data Transparency:6 estimated
Long-Context Retrieval
3/5
RULER
no data
NIAH
no data
LooGLE
77.550pEstimated
LongBench
603pEstimated
∞Bench
80.372pEstimated
Multi-Turn Recall
1/2
LoCoMo
68.923pEstimated
MemoryBank
no data
Cross-Session Memory
1/1
LongMemEval
70.624pEstimated
Multi-Hop QA
1/3
BABILong
80.390pEstimated
MultiHop-RAG
no data
HotpotQA
no data
Agent Task Memory
0/1
AgentBench-Mem
no data
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:Arena estimate — derived from capability profile, not independently verified