Self-Hosted Memoryfor AI Agents

Token Intelligence

NovaCortex retrieves only what the agent needs right now — no full-context dumps. Salience decay model learns what to forget. Typical savings: 85–95% fewer input tokens.

Graph-Native Memory

Four typed memory kinds (Episodic, Semantic, Procedural, Working) stored in a force-directed relation graph. Not a flat vector store — a real knowledge graph.

PMF — Portable Memory Format

Open RFC-001 standard for vendor-neutral AI memory interchange. Export everything to a single PMF file, import into any compatible system. Merkle-verified. No lock-in.

Knowledge Buckets

Organize memories into named buckets with per-agent access control. Fine-grained isolation for multi-agent systems. Namespace federation across deployments.

MCP Built-In

Native Model Context Protocol server — zero configuration. Works with Claude Desktop, Cursor, Windsurf, and any MCP-compatible client out of the box.

Free Self-Hosting

docker compose up. Runs on any VPS, Unraid, Raspberry Pi, or bare metal. Full stack — API, web UI, vector DB, graph DB, cache — in one command.

{"pmf"   "1.0","namespace"   "agent-research","exported"   "2026-04-12T09:00:00Z","memories"   [    {"id"       "m_01","content"       "HNSW index achieves sub-ms recall at 1M+ vectors","embedding"       [0.021, -0.134, 0.892, ...],"tags"       ["research", "performance"]    }  ]}

# Pull and start
docker compose up -d
# Web:  http://localhost:3000
# API:  http://localhost:3001
# Docs: http://localhost:3000/docs

What is NovaCortex?

NovaCortex is an open-source, self-hostable memory layer for AI agents. It gives agents persistent, queryable memory — typed memories linked by a relation graph, with semantic and text retrieval — exposed through a native MCP server, a REST API, TypeScript/Python SDKs, and a CLI. It runs entirely on your own infrastructure via Docker.

Does my data leave my infrastructure?

No. NovaCortex is self-hosted, so your knowledge base lives on your own servers and never goes to a third-party cloud. Retrieval also sends only the few relevant snippets to the model per query, instead of your whole knowledge base — so even less of your data is exposed than with context-dumping approaches.

How is NovaCortex different from Mem0 and Zep?

Mem0 and Zep are cloud-first. NovaCortex is privacy-first and self-hostable: your data stays on your infrastructure, and a portable open format (PMF) lets you export the entire memory graph at any time with zero lock-in. It is also MCP-native, so it plugs directly into Claude and Cursor.

How much does NovaCortex cost?

The full engine is free and open source (Apache-2.0) to self-host, with 3 namespaces. Pro is a one-time $10 license that unlocks 10 namespaces, namespace federation, and higher limits while staying self-hosted. Enterprise is custom, with unlimited namespaces and an SLA.

How much does retrieval save versus putting everything in context?

In a reproducible benchmark (15 questions, gpt-4o-mini, top-5 retrieval), retrieving only the relevant memories used 85% to 98.8% fewer input tokens than dumping the whole knowledge base into the prompt — about 65x cheaper per query on a 26K-token knowledge base — at parity answer accuracy and 100% retrieval recall. Dump cost grows with the knowledge base; retrieval stays flat.

Does NovaCortex work with Claude and Cursor?

Yes. NovaCortex ships a native Model Context Protocol (MCP) server with 15 tools, so Claude, Cursor, or any MCP-compatible agent can store and retrieve memory directly. You can also use the REST API, the TypeScript or Python SDK, or the CLI.

Can I export my memory, or am I locked in?

You can export everything. The Portable Memory Format (PMF) exports the full graph — memories, relations, and embeddings — as JSON, compact binary (MessagePack), or AES-256-GCM-encrypted, all with integrity hashes. There is no lock-in.