Back to Arena

LongMem

by UCSB / Microsoft Research

System Card

OrganizationUCSB / Microsoft Research
Released2023-06
Architectureexternal-memory-network / Decoupled frozen encoder + residual side network
DetailsDecoupled network with a frozen LLM backbone as the memory encoder and a trainable residual side network as retriever/reader. Caches up to 65k tokens in a non-differentiable memory bank, mitigating memory staleness.
Parameters—
Domainlong-contextlifelong-learning
Open SourceYes
PaperView Paper
CodeRepository
neurips-2023side-networkin-context-learningmemory-bank

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Data Transparency:6 estimated
Long-Context Retrieval
3/5
RULER
no data
NIAH
no data
LooGLE
73.123pEstimated
LongBench
603pEstimated
∞Bench
78.750pEstimated
Multi-Turn Recall
1/2
LoCoMo
65.818pEstimated
MemoryBank
no data
Cross-Session Memory
1/1
LongMemEval
65.920pEstimated
Multi-Hop QA
1/3
BABILong
72.624pEstimated
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