Back to Arena

Mnemosyne

by independent

System Card

Organizationindependent
Released2025-10
Architecturegraph-rag / Human-inspired probabilistic recall with decay
DetailsUnsupervised long-term memory for edge LLMs using graph-structured storage, modular substance/redundancy filters, memory pruning, and probabilistic recall with temporal decay modeled on human memory. Includes a "core summary" of personality/domain details.
Parameters—
Domainpersonalizationepisodic-sessionagent-memory
Open SourceNo
PaperView Paper
edge-llmdecaygraphunsupervisedpersonalized

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
54.51p
MemoryBank
64.413p
Cross-Session Memory
1/1
LongMemEval
59.912p
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
66.128p
HotpotQA
64.830p
Agent Task Memory
1/1
AgentBench-Mem
617p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:arXiv:2510.08601 Table 2 — Overall J-score, Llama3.1-8B-Instruct; SingleHop 62.78, MultiHop 49.53, OpenDomain 60.42, Temporal 53.03Mnemosyne paper (arXiv:2510.08601); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)Mnemosyne paper (arXiv:2510.08601); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)Mnemosyne paper (arXiv:2510.08601); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)Mnemosyne paper (arXiv:2510.08601); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)Mnemosyne paper (arXiv:2510.08601); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)