memgraph-agent — Graph Memory for AI Agents

Three pillars of intelligent memory

Inspired by SPRIG, built for production.

NER Entity Graph

CPU-only entity extraction via spaCy + domain dictionaries. Supports English & Chinese. Custom dictionaries for finance, crypto, and AI terms. No LLM calls, no API keys.

spaCy (15MB) + regex → <1ms extraction

PPR Multi-hop Retrieval

Personalized PageRank walks the co-occurrence graph to find entities N hops away. Search "BTC" → find "Freqtrade" (2 hops). With explainable path traces.

α=0.85 damping • 1.9ms avg • 2-hop default

Hybrid Search

Fuses vector similarity (ChromaDB + sentence-transformers) with graph PPR scores. 28% higher multi-hop recall than vector-only. Tunable α blend factor.

score = α·vector + (1-α)·PPR

Architecture

📝

Text Input

MEMORY.md
Daily notes
Live ingest

🔬

NER Extract

spaCy + regex
EN/ZH support
Custom dicts

🕸️

Graph + Vector

NetworkX graph
ChromaDB vectors
Co-occurrence

🔍

Hybrid Search

PPR multi-hop
Vector similarity
Fused ranking

Quick Start

1. Install

pip install memgraph-agent
python -m spacy download en_core_web_sm

2. Build your knowledge graph

from memgraph_agent import MemoryGraph

mg = MemoryGraph()
result = mg.build()
# → 102 nodes, 728 edges, 8 communities

3. Search with multi-hop retrieval

# Hybrid search (vector + graph PPR)
results = mg.search("BTC trading strategy")

# Pure graph search (PPR only, ~1.9ms)
results = mg.ppr_search("BTC")

# Add new memories
mg.ingest("CrewAI v0.80 released with async support")

4. Visualize & explore

# Interactive HTML visualization (pyvis)
mg.visualize()

# Detect entity communities (Louvain)
communities = mg.communities()

# Export for external tools
mg.export()  # → GraphML format

How it compares

System	Monthly Cost	Latency	Multi-hop	CPU-only
GPT-4 extraction	~$450	2-5s	Prompt-dependent	✗
Mem0	~$200-500	1-3s	Limited	✗
GraphRAG	~$100-300	Minutes	✓	✗
memgraph-agent	$0	<2ms	✓ PPR	✓