Quantum Memory Graph (QMG)

Quantum-enhanced long-term memory for AI agents. Uses chunked gte-large embeddings for state-of-the-art semantic retrieval with QAOA+CVaR quantum subgraph optimization for graph-based reasoning.

Features

#1 on LongMemEval — 98.6% R@5, 99.4% R@10, 0.9426 NDCG
Chunked retrieval — 500-char blocks with 100-char overlap, mean-of-top-3 per session
QAOA+CVaR optimization — 12.8% edge over greedy on graph/PCE tasks
GPU accelerated — runs on NVIDIA GB10 (DGX Spark)
Cascade recall — personal graph → historical archive fallback
Per-agent isolation — each agent gets their own isolated memory graph

How It Works

Session chunking — Conversations split into overlapping 500-char chunks
Embedding — Chunks encoded with gte-large (1024-dim sentence transformer)
Scoring — Per-session score = mean of top-3 chunk cosine similarities
Refinement — Top-N candidates optionally refined via QAOA+CVaR subgraph optimizer
Cascade — Personal graph first, historical archive fallback if relevance < 0.4

Performance

| Metric | Score | |--------|:-----:| | R@1 | 90.6% | | R@5 | 98.6% | | R@10 | 99.4% | | NDCG@10 | 0.9426 |

Benchmark: LongMemEval-S (500 questions, 18,464 sessions), May 28 2026

Usage

from quantum_memory_graph import MemoryGraph

mg = MemoryGraph()
mg.store("Project Alpha uses React frontend with TypeScript.")
mg.store("Project Alpha backend is FastAPI with PostgreSQL.")

# Recall — chunked semantic retrieval + optional QAOA refinement
results = mg.retrieve("What is Project Alpha's tech stack?", top_k=5)

Or as a FastAPI server for agent integration:

pip install quantum-memory-graph
quantum-memory-graph serve  # serves /store, /recall, /stats endpoints

Requirements

Python 3.10+
sentence-transformers
numpy
(optional) qiskit for QAOA optimization on real hardware

License

MIT