Back to Arena

MemoryBank

by Harbin Institute of Technology / SenseTime

System Card

OrganizationHarbin Institute of Technology / SenseTime
Released2023-05
Architecturevector-rag / Ebbinghaus forgetting curve + vector store
DetailsVector-stored episodic memories with a novel forgetting/reinforcement mechanism inspired by the Ebbinghaus forgetting curve. Memories decay and strengthen based on elapsed time and relevance, with periodic personality and event summarization.
Parameters—
Domainpersonalizationepisodic-sessionagent-memory
Open SourceYes
PaperView Paper
CodeRepository
aaai-2024ebbinghausdecaycompanionbenchmark

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
59.111p
MemoryBank
81.888p
Cross-Session Memory
1/1
LongMemEval
57.55p
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
63.617p
HotpotQA
62.422p
Agent Task Memory
1/1
AgentBench-Mem
6514p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:arXiv:2305.10250 Table 2 — SiliconFriend-ChatGPT English Ranking — same paper as memorybank entryMemoryBank paper (arXiv:2305.10250); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)MemoryBank paper (arXiv:2305.10250); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)MemoryBank paper (arXiv:2305.10250); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)MemoryBank paper (arXiv:2305.10250); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)MemoryBank paper (arXiv:2305.10250); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)