Back to Arena

MIRIX

by MIRIX AI (Wang, Chen)

System Card

OrganizationMIRIX AI (Wang, Chen)
Released2025-07
Architectureagentic-workflow / Six-type memory + multi-agent coordinator
DetailsSix memory types (Core, Episodic, Semantic, Procedural, Resource, Knowledge Vault) coordinated by a multi-agent framework that dynamically schedules reads/writes. Supports text + visual modalities.
Parameters—
Domainagent-memorypersonalization
Open SourceYes
PaperView Paper
WebsiteVisit
CodeRepository
multi-agentmultimodalsix-memory-typesscreenshot-vqa

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
85.496p
MemoryBank
79.475p
Cross-Session Memory
1/1
LongMemEval
81.387p
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
7579p
HotpotQA
79.996p
Agent Task Memory
1/1
AgentBench-Mem
7226p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:arXiv:2507.07957 results table — Overall accuracy; SingleHop 85.11, MultiHop 83.70, OpenDomain 65.62, Temporal 88.39MIRIX paper (arXiv:2507.07957); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)MIRIX paper (arXiv:2507.07957); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)MIRIX paper (arXiv:2507.07957); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)MIRIX paper (arXiv:2507.07957); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)MIRIX paper (arXiv:2507.07957); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)