Supermemory

by Supermemory

System Card

OrganizationSupermemory

Released2024-06

Architecturegraph-rag / Vector graph engine with ontology-aware edges

DetailsCustom vector graph engine that tracks relationships between memories instead of similarity scores. Five-layer context stack: connectors, extractors, retrieval, memory graph, and user profiles. Handles contradictions, temporal changes, and knowledge updates without corrupting existing state. Sub-300ms recall speed. Shipped with MCP plugin for cross-LLM memory portability across ChatGPT, Claude, and Gemini.

Parameters—

Domainagent-memorypersonalizationepisodic-sessionknowledge-graph

Open SourcePartial

PaperView Paper

WebsiteVisit

vector-graphmcpcross-llmsub-300msknowledge-updatestemporal-reasoningagent-memoryyc-backed

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

83.594p

81.284p

Cross-Session Memory

1/1

85.497p

Multi-Hop QA

2/3

BABILong

no data

72.454p

78.693p

Agent Task Memory

1/1

68.518p

Personalization

0/1

PerLTQA

no data

Factuality / Grounding

0/1

RAGAS

no data

Sources:Supermemory research — LongMemEval benchmark (overall accuracy 85.4%, single-session retrieval 92.3%, knowledge updates 89.7%, temporal reasoning 82.0%)Supermemory research — LoCoMo Recall@10 83.5%, P@1 59.7%, NDCG@10 71.1%Supermemory (supermemoryai/supermemory); evaluated on HotpotQA (Stanford / CMU, 2018)Supermemory (supermemoryai/supermemory); evaluated on MemoryBank (Sun Yat-sen University, 2023)Supermemory (supermemoryai/memorybench); evaluated on MultiHop-RAG (HKUST, 2024)Supermemory (supermemoryai/memorybench); evaluated on AgentBench Memory Track (Tsinghua KEG, 2023)