MemVault

A Memory Server for AI Agents. Runs on Postgres + pgvector. Now supporting 100% Local/Offline execution via Ollama.

I got tired of setting up Pinecone/Weaviate and writing the same embedding boilerplate for every small AI agent I built.

I wanted something that:

1. Just runs on PostgreSQL (which I already use).
2. Handles the chunking & embedding automatically.
3. Lets me visualize the retrieval process (because debugging vector similarity in JSON logs is difficult).
4. Can run offline without API bills.

So I built MemVault. It is a Node.js wrapper around pgvector with a generic Hybrid Search engine.

---

Quick Start: Choose your setup

You can run this entirely on your own machine (Docker), or use the managed API to skip the server maintenance.

Feature	Self-Hosted (Docker)	Managed API (RapidAPI)
Price	Free (Open Source)	Free Tier available
Embeddings	Ollama (Local) or OpenAI	OpenAI (Managed)
Setup Time	~15 mins	30 seconds
Data Privacy	100% on your server	Hosted by us
Maintenance	You manage updates/uptime	We handle everything
Link	Scroll down to Docker	Get API Key

---

Hybrid Search 2.0 (The Algorithm)

Most RAG pipelines only use Vector Search. MemVault uses a 3-way weighted score to find the most relevant context:

1. Semantic (Vector): Uses Cosine Similarity via pgvector to understand meaning.
2. Exact Match (Keyword): Uses BM25 (Postgres tsvector) to find exact product IDs or error codes that vectors miss.
3. Recency (Time): A decay function prioritizing recent memories.

FinalScore = (Vector * 0.5) + (Keyword * 0.3) + (Recency * 0.2)

---

The Visualizer

The hardest part of RAG is knowing why your bot retrieved specific context. MemVault comes with a dashboard to visualize the vector search in real-time.

(Live Demo: memvault-demo.vercel.app)

---

Installation (NPM SDK)

Whether you self-host or use the Cloud API, the SDK works the same way.

npm install memvault-sdk-jakops88

import { MemVault } from 'memvault-sdk-jakops88';

// Point to local instance or RapidAPI
const memory = new MemVault({
  apiKey: "YOUR_KEY", 
  baseUrl: "http://localhost:3000" 
});

// 1. Store a memory (Auto-embedding via Ollama/OpenAI)
await memory.store({
  sessionId: "user-123",
  text: "The user prefers strictly typed languages like TypeScript.",
  importanceHint: "high"
});

// 2. Retrieve relevant context (Hybrid Search)
const result = await memory.retrieve({
  sessionId: "user-123",
  query: "What tech stack should I recommend?",
  limit: 3
});

---

Self-Hosting (Docker)

You can run the entire stack (API + DB + Embeddings) offline.

Prerequisites

• Docker & Docker Compose
• Ollama (optional, for local embeddings)

1. Clone the repository

git clone https://github.com/jakops88-hub/Long-Term-Memory-API.git
cd Long-Term-Memory-API

2. Configure Environment

cp .env.example .env

To use local embeddings (free/offline), set the provider to ollama in your .env file:

EMBEDDING_PROVIDER=ollama
OLLAMA_BASE_URL=http://host.docker.internal:11434/api
OLLAMA_MODEL=nomic-embed-text

Ensure you have pulled the model in Ollama: ollama pull nomic-embed-text

3. Start the stack

docker-compose up -d

The API is now available at http://localhost:3000.

---

Architecture

• Runtime: Node.js & TypeScript
• Database: PostgreSQL + pgvector
• Search: Hybrid (Vector + BM25 Keyword Search)
• ORM: Prisma
• Visualization: React + react-force-graph-2d

Contributing

This is a side project that grew into a tool. Issues and PRs are welcome. Specifically looking for help with:

• Metadata Filters: Adding structured filtering alongside vectors.
• Security: Implementing session-level encryption.

License

MIT