Back to Arena

D-Mem

by You et al. (2025)

System Card

OrganizationYou et al. (2025)
Released2025-11
Architecturehybrid / System-1 vector retrieval + System-2 full deliberation fallback
DetailsDual-process memory: System-1 uses fast vector retrieval for routine queries (Mem0-style); System-2 invokes a Full Deliberation module as a high-fidelity fallback when fine-grained context is needed.
Parameters—
Domainagent-memoryepisodic-session
Open SourceNo
PaperView Paper
dual-processsystem-1-2fallbackmem0-compatible

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
78.478p
MemoryBank
77.965p
Cross-Session Memory
1/1
LongMemEval
77.359p
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
7579p
HotpotQA
79.494p
Agent Task Memory
1/1
AgentBench-Mem
7226p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:arXiv:2603.18631 Table 1 — LLM-judge with GPT-4o-mini Full Deliberation; F1 55.3, BLEU 44.2. Using LLM-judge as primary to match Mem0/Backboard/MemR3 comparabilityD-Mem paper (arXiv:2603.18631); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)D-Mem paper (arXiv:2603.18631); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)D-Mem paper (arXiv:2603.18631); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)D-Mem paper (arXiv:2603.18631); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)D-Mem paper (arXiv:2603.18631); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)