Back to Arena

Larimar

by IBM Research

System Card

OrganizationIBM Research
Released2024-03
Architectureexternal-memory-network / Brain-inspired distributed episodic memory
DetailsAttaches a distributed episodic memory module to an LLM for one-shot fact editing without retraining. Supports selective forgetting, leakage prevention, and context-length generalization.
Parameters—
Domainknowledge-graphlifelong-learning
Open SourceYes
PaperView Paper
WebsiteVisit
CodeRepository
icml-2024fact-editingepisodicibmmodel-agnostic

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
66.419p
MemoryBank
60.46p
Cross-Session Memory
1/1
LongMemEval
68.722p
Multi-Hop QA
3/3
BABILong
75.755p
MultiHop-RAG
71.345p
HotpotQA
69.943p
Agent Task Memory
0/1
AgentBench-Mem
no data
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:Larimar paper (arXiv:2403.11901); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)Larimar paper (arXiv:2403.11901); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)Larimar paper (arXiv:2403.11901); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)Larimar paper (arXiv:2403.11901); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)Larimar paper (arXiv:2403.11901); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)Larimar paper (arXiv:2403.11901); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)