Mnemosyne

by Johns Hopkins / independent (2025)

System Card

OrganizationJohns Hopkins / independent (2025)

Released2025-10

Architecturegraph-rag / Graph memory with human-inspired decay, refresh, and core-summary

DetailsGraph-structured LTM with modular substance/redundancy filters, commit and pruning, probabilistic recall using temporal decay and refresh. Adds a concentrated "core summary" derived from a fixed-length memory-graph subset.

Parameters—

Domainpersonalizationlifelong-learningepisodic-session

Open SourceNo

PaperView Paper

edgedecay-refreshgraphunsupervised

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

62.610p

Cross-Session Memory

1/1

LongMemEval

61.415p

Multi-Hop QA

1/3

BABILong

73.738p

MultiHop-RAG

no data

HotpotQA

no data

Agent Task Memory

1/1

AgentBench-Mem

61.48p

Personalization

1/1

PerLTQA

67.79p

Factuality / Grounding

0/1

RAGAS

no data

Sources:arXiv:2510.08601 Table 2 — Same paper as mnemosyne — paper is titled 'for Edge-Based LLMs'; edge is the main design target Mnemosyne 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 MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)Mnemosyne paper (arXiv:2510.08601); evaluated on PerLTQA: A Personal Long-Term Memory Question Answering Dataset (PolyU, 2402)Mnemosyne paper (arXiv:2510.08601); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)Mnemosyne paper (arXiv:2510.08601); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)