Back to Arena

MemoryLLM

by UCSD / Apple (Wang et al.)

System Card

OrganizationUCSD / Apple (Wang et al.)
Released2024-02
Architectureexternal-memory-network / Fixed-size latent memory pool updated in-place
DetailsLLM with an integrated fixed-size memory pool in latent space that self-updates as new text is injected. Parameters within the pool are continuously overwritten, enabling knowledge integration without retraining.
Parameters—
Domainlifelong-learninglong-context
Open SourceYes
PaperView Paper
CodeRepository
icml-2024self-updatablelatent-poolcontinual

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Long-Context Retrieval
3/5
RULER
no data
NIAH
no data
LooGLE
79.282p
LongBench
603p
∞Bench
8069p
Multi-Turn Recall
1/2
LoCoMo
76.362p
MemoryBank
no data
Cross-Session Memory
1/1
LongMemEval
75.248p
Multi-Hop QA
1/3
BABILong
82.194p
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:MemoryLLM paper (arXiv:2402.04624); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)MemoryLLM paper (arXiv:2402.04624); evaluated on InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens (Tsinghua / OpenBMB, 2402)MemoryLLM paper (arXiv:2402.04624); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)MemoryLLM paper (arXiv:2402.04624); evaluated on LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (Tsinghua KEG, 2308)MemoryLLM paper (arXiv:2402.04624); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)MemoryLLM paper (arXiv:2402.04624); evaluated on LooGLE: Can Long-Context Language Models Understand Long Contexts? (Peking University, 2311)