Back to Arena

MemOS

by MemTensor (Li, Zhang, et al.)

System Card

OrganizationMemTensor (Li, Zhang, et al.)
Released2025-05
Architectureagentic-workflow / Memory OS with MemCube units across plaintext/activation/parameter
DetailsTreats memory as a managed system resource. MemCube units encapsulate content plus metadata and can be composed/migrated/fused across plaintext, activation, and parameter memory. Implements permissioning, lifecycle management, and evolvability.
Parameters—
Domainagent-memorylifelong-learning
Open SourceYes
PaperView Paper
CodeRepository
memory-osmemcubesystem-resourcelifecycle

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
75.857p
MemoryBank
75.349p
Cross-Session Memory
1/1
LongMemEval
77.863p
Multi-Hop QA
2/3
BABILong
72.117p
MultiHop-RAG
no data
HotpotQA
71.852p
Agent Task Memory
1/1
AgentBench-Mem
7226p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:arXiv:2507.03724 — MemOS-1031 average, Table 3 — Average across LongMemEval categories; outperforms Memobase 72.4%arXiv:2507.03724 — MemOS-1031 overall LLM Judge Score — Headline LoCoMo LLM-as-Judge score; also shown on MemTensor GitHub READMEMemOS paper (arXiv:2505.22101); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)MemOS paper (arXiv:2505.22101); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)MemOS paper (arXiv:2505.22101); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)MemOS paper (arXiv:2505.22101); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)