Open Brain System The open-source AI-integrated brain system — pgvector + MCP + Supabase
2026-04-21 / Guerin Green / Open Brain System

Open Brain System: The Reference FAQ

Forty-plus questions and direct answers on open brain architecture, MCP integration, pgvector operations, and how it compares to everything else.

A Short Preamble

This resource serves as a centralized open brain system FAQ, consolidating the most frequent technical inquiries found across the site's deep-dive architectural guides. Each response provides a direct answer to accelerate implementation and troubleshooting for AI memory systems.

For comprehensive analysis on specific implementations—such as integrating pgvector with Mem0 or deploying local MCP servers—links to full articles are provided within each section. These guides cover the shift toward agentic, persistent memory and the hardware constraints affecting HBM and DDR5 supply in 2026.

If a specific technical challenge or edge case is missing from this list, contact the editor directly at guerin@novcog.com to request an addition to the open brain system FAQ database.

Questions answered

What readers usually ask next.

What is an open brain system?
In the context of AI, an 'open brain' refers to OpenMemory, a privacy-focused, self-contained memory layer for AI agents. Unlike cloud-based memory, it allows for local hosting and data portability via MCP servers, ensuring your agentic recall remains private and under your control.
What is the Model Context Protocol (MCP)?
MCP is a standardized protocol that enables AI applications to connect seamlessly to external data sources and memory layers. It allows tools like OpenMemory to act as a local server, providing real-time context to LLMs without requiring constant API calls or data egress.
Why should I care about open brain architecture?
Open brain architecture solves the 'amnesia' problem in AI agents by providing persistent, long-term recall across multiple sessions. By using a local or self-hosted layer, you avoid vendor lock-in and maintain strict data privacy while improving agent performance on LOCOMO benchmarks.
How is an open brain different from a second brain?
A 'second brain' (like Tiago Forte's BASB) is a manual system for organizing notes and knowledge. An open brain is an automated, machine-readable memory layer that uses vector embeddings to allow AI agents to retrieve and apply information autonomously in real-time.
Can Obsidian be used as an open brain?
Obsidian serves as a static knowledge base, but it becomes part of an open brain system when connected via an MCP server or RAG pipeline. By indexing your Markdown files into a vector store like pgvector, your AI agents can query your notes as active memory.
Is Tiago Forte's methodology considered an open brain?
No, the PARA method is a human-centric organizational framework for digital filing. An open brain refers to the technical infrastructure—specifically vector databases and agentic frameworks like Mem0—that allows AI to programmatically recall information.
What is pgvector?
pgvector is an open-source extension for PostgreSQL that enables the storage and querying of vector embeddings. It is widely used as the backend for frameworks like LangChain and Mem0 to perform similarity searches for AI agent memory.
What's the cheapest way to build an open brain?
The most cost-effective approach is self-hosting a PostgreSQL instance with pgvector on existing hardware or using Supabase's free tier. Pairing this with local embedding models and an MCP server eliminates monthly subscription fees associated with proprietary memory clouds.
How long does it take to build an open brain system?
A basic implementation using a pre-built MCP server and a managed vector database can be set up in a few hours. However, optimizing for low latency and high recall accuracy—especially for voice agents—typically requires several days of tuning.
Do I need to know Python or SQL to build an open brain?
While no-code tools are emerging, basic knowledge of Python is essential for configuring frameworks like Mem0 or LangChain. SQL is necessary if you are managing your own pgvector instance to optimize query performance and index management.
Is Supermemory considered an open brain?
Supermemory functions as a memory layer, but the term 'open brain' specifically emphasizes the portability and local-first nature of systems like OpenMemory. The key distinction is whether the data remains in a proprietary silo or is exportable via open protocols.
What is NovCog Brain?
NovCog Brain is an emerging approach to cognitive architecture for AI, focusing on how agents structure memory over time. It aims to move beyond simple vector retrieval toward more complex, hierarchical recall systems.
Can I use an open brain system with Cursor?
Yes, by utilizing MCP-compatible extensions or providing the AI with access to a local vector index of your documentation. This allows Cursor to have long-term context of your entire codebase and architectural decisions across different projects.
Can I use an open brain with Claude Desktop?
Yes, Claude Desktop supports MCP servers. By running a local OpenMemory or custom MCP server, you can give Claude the ability to read and write to your personal persistent memory store directly from the desktop app.
What embedding model should I use for an open brain?
For most users, OpenAI's text-embedding-3-small offers a great balance of cost and performance. For those prioritizing privacy and local execution, BGE-M3 or Cohere's multilingual models are top choices for high-dimensional retrieval.
Is IVFFlat or HNSW the right pgvector index?
HNSW is generally preferred for open brain systems because it offers faster query speeds and higher recall accuracy. IVFFlat is more memory-efficient and faster to build, making it better for massive datasets where slight precision loss is acceptable.
How do I migrate from Obsidian to an open brain?
The process involves exporting your Markdown files and passing them through an embedding pipeline. You can use a tool like LlamaIndex or a custom Python script to chunk the text, generate embeddings, and upload them into pgvector.
How do I migrate from Supermemory to a self-hosted open brain?
You must export your stored data in JSON or CSV format, then re-embed that data using your chosen local model. Finally, import the resulting vectors into your own pgvector instance and connect it via an MCP server.
What is the acceptable latency budget for an open brain?
For text agents, retrieval should happen in under 200-500ms to avoid breaking conversation flow. For real-time voice agents (e.g., ElevenLabs), the budget is much tighter, requiring optimized HNSW indexes and local hosting to minimize round-trip time.
How does an open brain scale past 1 million entries?
Scaling requires implementing partitioning in PostgreSQL and utilizing HNSW indexing to maintain search speed. For extremely large datasets, you may need to move from a single instance to a distributed vector database or implement hierarchical caching.
Does Supabase's free tier handle an open brain system?
Yes, for small to medium personal knowledge bases. The free tier includes pgvector support, which is sufficient for thousands of embeddings, though you may hit storage limits as your memory grows into the millions of tokens.
Is my data safe in an open brain system?
Safety depends on deployment. A local-first OpenMemory setup via MCP is highly secure because data never leaves your machine. If using a cloud provider like Supabase, security relies on their encryption and your API key management.
Can I share my open brain with other people?
Yes, if you use a centralized vector store like pgvector, you can grant access to other agents or users via API keys. For local systems, you would need to export your database dump and share the snapshot.
Is there a complete starter kit for building an open brain?
While no single 'box' exists, the combination of Mem0 for memory management, pgvector for storage, and an MCP server for connectivity serves as the current industry-standard blueprint.
What's the difference between an open brain and a knowledge graph?
An open brain typically uses vector embeddings for semantic similarity (finding things that 'feel' related). A knowledge graph uses explicit nodes and edges to define exact relationships, though hybrid systems now combine both for better recall.
Can I use multiple embedding models in one open brain?
Technically yes, but you cannot query across different models because their vector spaces are incompatible. You would need to maintain separate indexes or re-embed all data whenever you switch models.
Which AI clients currently support MCP?
Claude Desktop is the primary first-party client supporting the Model Context Protocol. Many open-source wrappers and IDEs like Cursor are rapidly integrating MCP to allow agents to access local memory servers.
How often should I re-embed my memories?
You only need to re-embed when you change your embedding model or significantly update the underlying data. If you upgrade from a legacy model to a newer one (e.g., BGE-M3), you must re-process all existing entries to maintain search accuracy.
Does it matter what Postgres version I use for an open brain?
Yes, you must use a version that supports the pgvector extension. It is highly recommended to use PostgreSQL 15 or 16 to take advantage of improved indexing performance and better memory management.
Can I run an open brain system on a Raspberry Pi?
Yes, provided you use lightweight local embedding models (like Sentence-Transformers) and a lean Postgres installation. However, retrieval latency will be significantly higher than on x86 hardware with NVMe storage.