Back to Arena

Titans

by lucidrains (community) / paper by Google Research

System Card

Organizationlucidrains (community) / paper by Google Research
Released2025-01
Architectureexternal-memory-network / Neural memory module (Memory-as-Context)
DetailsUnofficial reference PyTorch implementation of Google's Titans architecture: a learned NeuralMemory module that updates parameters at test time, combined with Memory-As-Context (MAC) transformer wrapping for long-term memory.
Parameters—
Domainlong-contextlifelong-learning
Open SourceYes
PaperView Paper
CodeRepository
neural-memorytest-time-learningtransformergoogle-research

Capability Profile

Benchmark Scores

8 of 14 benchmarks
Data Transparency:2 self-reported6 estimated
Long-Context Retrieval
5/5
RULER
97.999pSelf-Reported
NIAH
98.892pSelf-Reported
LooGLE
81.395pEstimated
LongBench
603pEstimated
∞Bench
87.494pEstimated
Multi-Turn Recall
1/2
LoCoMo
68.622pEstimated
MemoryBank
no data
Cross-Session Memory
1/1
LongMemEval
7545pEstimated
Multi-Hop QA
1/3
BABILong
82.496pEstimated
MultiHop-RAG
no data
HotpotQA
no data
Agent Task Memory
0/1
AgentBench-Mem
no data
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:arXiv:2501.00663 Table 2 — Titans (MAC) avg on RULER S-NIAH-PK/N/W at 2K/4K/8K/16KarXiv:2501.00663 Table 2 — Titans (MAC) S-NIAH-PK avg: 99.2/98.8/99.0/98.4 at 2K/4K/8K/16KArena estimate — derived from capability profile, not independently verified