// WHAT CORTEX GIVES YOU

All-in-one Memory Infrastructure.

Server-side ONNX embeddings

Embeddings run locally inside Harper with no external API calls. No OpenAI key, no Voyage key, nothing to configure. Text is embedded automatically when stored and searched automatically when queried.

HNSW vector indexing

384-dimensional cosine similarity search built into the table layer. Sub-millisecond recall on millions of memories.

Multi-agent isolation

Namespace memories by agent ID. Shared table, scoped reads. Or give each agent its own table. Both patterns, one config.

Real-time propagation (MQTT)

When one agent learns something, every subscribed agent gets it immediately. Built-in messaging. No Kafka. No Redis Pub/Sub.

Semantic + attribute search

Hybrid retrieval. Vector similarity for meaning, structured filters for classification, source, channel, and author. One query, both axes.

Distributed via Harper Fabric

Host locally for free or deploy across a secure distributed cloud cluster. Memory survives restarts, scales horizontally, and syncs across regions for low latency access wherever systems need it.

// ARCHITECTURE

How it works.

// WHY CORTEX

Compared to the alternatives.

CORTEX

Mem0

LanceDB

Markdown Files

Local

Single runtime, $0 embeddings

npm install, run locally. ONNX embeddings built in.

Docker stack + external APIs

3 containers (API + Postgres + Neo4j) + OpenAI key

Embedded file + external API

Local vector store, but needs OpenAI for embeddings

Flat file, no dependencies

No vector search, no semantic retrieval

Cloud
(Free Tier)

Up to 1.6M retrievals/mo

2M writes. 5GB storage. 20M real-time messages. $0 embedding cost.

×1,600 more vs. Mem0

1k retrievals/mo

10K memories. Plus external embedding API costs.

No cloud option

Cloud
(@ ~$250)

Up to 33M retrievals/mo

20M writes. 10GB storage. 500M real-time messages. $0 embedding cost.

×660 more vs. Mem0

50k retrievals/mo

$249/month, fully managed

No cloud option

Persistence

Distributed via Fabric

Survives restarts, replicates to multiple locations

Cloud or self-hosted

Platform managed, or Docker (API + Postgres + Neo4j)

Local file

Dies with the agent process

Local file

Multi-agent sharing

Namespace isolation + shared pools

Per-agent, per-team, or global scoping

User/agent/session scoping

Supports user_id, agent_id, and run_id

Known memory bleed issues

HNSW vector + attribute filters

Hybrid retrieval in a single query

Semantic + graph relationships

Entity extraction via Neo4j

HNSW vector search

Text match only

Also an app runtime

✔ Unified Runtime (Harper)

DB, cache, API, messaging in one process

✕ Memory layer only

✕ Vector store only

// ECOSYSTEM

Plug into any agent stack.

Cortex connects to the tools and frameworks your agents already use.

LIVE

// GET STARTED

Three steps. Five minutes.

Create a Free Fabric Cluster

Cortex runs on Harper Fabric, a distributed application runtime with built-in vector search. Create a free cluster in under a minute. No credit card required.

Create Free Cluster

Deploy Cortex

Clone the Cortex repo, add your cluster credentials, and deploy. This sets up memory tables, vector indexes, and API endpoints on your Fabric cluster. Embeddings run server-side via ONNX, no external API key needed.

# 1. Clone the repo and install dependencies
‍

git

 clone https://github.com/HarperFast/Cortex.git

cd

 Cortex

npm install

 -g harperdb

npm install

‍

# 2. Add your Fabric cluster credentials

cp

 .env.example .env

# Edit .env → add cluster URL, username, and password
# Anthropic key optional (for memory classification)
# Embeddings run locally — no key needed

‍

# 3. Deploy to your cluster
‍

npm run deploy

Connect Your Tools

Cortex is running. Now install the package that matches how you work. An SDK for custom code, an MCP server for AI-assisted editors, or a plugin for agent frameworks.

SDK — for custom agents and backend services.

# 1. Install
npm i @harperfast/cortex-client

# 2. Store and search memories
import { CortexClient } from "@harperfast/cortex-client";

const cortex = new CortexClient({
url: "https://your-cluster.harperdbcloud.com",
token: "your-auth-token",
});

// Store a memory
await cortex.memory.store({
content: "Harper Fabric replicates writes across regions via MQTT, so Cortex memories stored in us-east are immediately available in eu-west without a separate sync layer",
source: "architecture-decision",
});

// Search by meaning — not keywords
const results = await cortex.memory.search({
query: "how does memory sync work across regions",
});

MCP Server — for Claude, Cursor, Windsurf

# 1. Install
‍

npm

 i @harperfast/cortex-mcp-server

# 2. See the setup guide for your editor
# → https://github.com/HarperFast/Cortex/blob/main/packages/cortex/docs/mcp-setup.md

OpenClaw Plugin — drop-in memory for OpenClaw / NemoClaw

# 1. Install
npm i @harperfast/openclaw-memory

# 2. See plugin docs for configuration and lifecycle hooks
# → https://github.com/HarperFast/Cortex/tree/main/packages/openclaw-memory

Explore Ecosystem

Shared Memory Layer for AI Agents

Agents Forget Everything

Context Rot

No Shared State

Local-Only Memory