mem0-dbay 0.1.0

PostgreSQL + pgvector graph store backend for Mem0 — powered by dbay.cloud

mem0-dbay

Drop Neo4j. Use one database for everything.

mem0-dbay is a graph store plugin for Mem0 that replaces Neo4j with dbay.cloud — a Serverless PostgreSQL platform with pgvector built-in. Your vector store and graph store run in one database, zero infrastructure to manage.

Why?

Self-hosting Mem0 requires 3 separate databases:

Mem0 → PostgreSQL + pgvector  (vector store)
     → Neo4j                  (graph store)    ← extra infra
     → SQLite                 (history)

Neo4j adds real pain:

• Another database to deploy, monitor, backup, and pay for
• 90-second cold start in Docker
• Completely different query language (Cypher)
• APOC plugin dependency

With mem0-dbay, everything runs on one dbay.cloud database:

Mem0 → dbay.cloud PostgreSQL  (vectors + graph + history)

One connection string. One database. Zero Cypher. Zero ops.

Quick start

1. Create a free database on dbay.cloud

Sign up at dbay.cloud and create a database. You'll get a connection string like:

postgres://user_xxx:[e-mail]:4432/my-mem0-db?sslmode=require&options=endpoint%3Dmy-mem0-db

pgvector is pre-installed. No extensions to enable, no tables to create — mem0-dbay handles everything automatically.

2. Install

pip install mem0-dbay

3. Use

import mem0_dbay

DBAY_URL = "postgres://user_xxx:[e-mail]:4432/my-mem0-db?sslmode=require&options=endpoint%3Dmy-mem0-db"

m = mem0_dbay.create_memory({
    "graph_store": {
        "provider": "dbay",
        "config": {"connection_string": DBAY_URL, "embedding_dimension": 1536}
    },
    "vector_store": {
        "provider": "pgvector",
        "config": {"connection_string": DBAY_URL, "embedding_model_dims": 1536}
    },
    "llm": {
        "provider": "openai",
        "config": {"api_key": "sk-..."}
    },
    "embedder": {
        "provider": "openai",
        "config": {"api_key": "sk-..."}
    },
})

# Add memories — entities and relationships are auto-extracted
m.add("Alice works at Google as an engineer", user_id="alice")
m.add("Bob manages Alice at Google", user_id="alice")

# Search
results = m.search("Where does Alice work?", user_id="alice")

That's it. No Neo4j to install. No Docker Compose with 3 containers. Just pip install and a connection string.

Works with any OpenAI-compatible LLM

Not just OpenAI — use DeepSeek, SiliconFlow, or any OpenAI-compatible provider:

m = mem0_dbay.create_memory({
    "graph_store": {
        "provider": "dbay",
        "config": {"connection_string": DBAY_URL, "embedding_dimension": 1024}
    },
    "vector_store": {
        "provider": "pgvector",
        "config": {"connection_string": DBAY_URL, "embedding_model_dims": 1024}
    },
    "llm": {
        "provider": "openai",
        "config": {
            "api_key": "sk-...",
            "openai_base_url": "https://api.siliconflow.cn/v1",
            "model": "deepseek-ai/DeepSeek-V3",
        }
    },
    "embedder": {
        "provider": "openai",
        "config": {
            "api_key": "sk-...",
            "openai_base_url": "https://api.siliconflow.cn/v1",
            "model": "BAAI/bge-m3",
        }
    },
})

Why dbay.cloud?

dbay.cloud is a Serverless PostgreSQL platform purpose-built for AI workloads.

	Self-hosted PG + Neo4j	dbay.cloud
Databases to manage	2 (PG + Neo4j)	0 — fully managed
Idle cost	24/7 running (~$50+/mo)	Scale-to-zero — free when idle
Cold start	Neo4j: 90s	~500ms warm / ~8s cold
Compute elasticity	Fixed size	1cu–8cu on demand — scales with your workload
pgvector	Install yourself	Pre-installed, HNSW indexes ready
BM25 full-text search	Not available	pg_search (ParadeDB) pre-installed
Database branching	Not possible	Git-style copy-on-write branches
Setup time	Hours (Docker, configs, health checks)	30 seconds

Scale-to-zero: pay nothing when idle

Most AI memory databases sit idle 90%+ of the time. On dbay.cloud, idle databases automatically suspend and cost nothing. When your agent makes a query, the database wakes up in ~500ms. Perfect for per-user memory databases where most users aren't active simultaneously.

Git-style branching: A/B test your memory strategies

Want to test a new reflection strategy without risking your production memories?

main (production memories)
  ├── branch: experiment-a  ← test new extraction prompts
  └── branch: experiment-b  ← test different similarity thresholds

Branches are instant copy-on-write — zero storage overhead. Compare results, keep the winner, delete the rest. No other memory infrastructure can do this.

Elastic compute: scale up for batch processing

Running a batch reflection job on thousands of memories? Scale up to 8cu (8 vCPU, 16 GB) for the job, then scale back to 1cu. No server resizing, no downtime.

How it works

Mem0's graph memory only does simple operations — no PageRank, no graph algorithms:

What Mem0 does	Neo4j (Cypher)	mem0-dbay (SQL)
Store entities	:__Entity__ nodes	graph_nodes table
Vector similarity	vector.similarity.cosine()	pgvector <=> operator
Store relationships	MERGE (a)-[r:TYPE]->(b)	INSERT INTO graph_edges
Find neighbors	MATCH (n)-[r]->(m)	JOIN graph_edges ON ...

Two PostgreSQL tables replace an entire Neo4j instance.

Also works with local PostgreSQL

If you prefer self-hosting, mem0-dbay works with any PostgreSQL that has pgvector:

"connection_string": "postgresql://user:pass@localhost:5432/mydb"

License

MIT