Back to Arena

MemoChat

by University of Warwick / Alibaba

System Card

OrganizationUniversity of Warwick / Alibaba
Released2023-08
Architecturehierarchical-summary / Memo-based iterative memorization-retrieval-response
DetailsTunes LLMs to maintain self-composed memos in an iterative "memorize-retrieve-respond" cycle. Instructions are reconstructed from public datasets so the model learns to write and query structured memos for long-range consistency.
Parameters—
Domainepisodic-sessionagent-memory
Open SourceYes
PaperView Paper
CodeRepository
instruction-tuningmemosfine-tunedlong-range

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Long-Context Retrieval
0/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
no data
∞Bench
no data
Multi-Turn Recall
2/2
LoCoMo
72.534p
MemoryBank
64.915p
Cross-Session Memory
1/1
LongMemEval
71.126p
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
67.930p
HotpotQA
67.638p
Agent Task Memory
1/1
AgentBench-Mem
70.422p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:MemoChat paper (arXiv:2308.08239); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)MemoChat paper (arXiv:2308.08239); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)MemoChat paper (arXiv:2308.08239); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)MemoChat paper (arXiv:2308.08239); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)MemoChat paper (arXiv:2308.08239); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)MemoChat paper (arXiv:2308.08239); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)