How to Build an AI Second Brain: The 4 Core Components
A practical framework for building a personal AI second brain using context, connections, cadence, and capabilities to run your work.
What is an AI second brain?
An AI second brain is a personal agent system, usually built on a coding agent like Claude Code plus a knowledge base of markdown files, that knows your work well enough to act like a teammate. It isn’t a chatbot you query occasionally. It’s a system wired into your calendar, email, task manager, and files that holds context on your business and helps you operate in it daily. The system rests on four components: context, connections, cadence, and capabilities.
TL;DR
- A second brain needs four components to work well: context, connections, cadence, and capabilities, a framework popularized by AI educator Nate Herk.
- Context means your agent should know your business “off the top of its head” without digging through files, achieved through core always-loaded documents plus a deeper searchable knowledge base.
- Connections determine whether your second brain can actually act where you work, including email, task managers, GitHub, Slack, and shared drives, typically wired in through MCP servers or small custom CLI tools.
- The simplest audit for context is asking a fresh session what your business does and who works there, and it should answer immediately like a teammate would.
- The simplest audit for connections is asking for a Monday morning catch-up and seeing if it can surface your real priorities across every platform without delay.
- You don’t need a sprawling tool stack: one builder runs an entire second brain on Claude Code, Obsidian, and a cheap VPS, because modern LLMs can handle diagramming, video editing, and presentation work directly through commands rather than bolted-on services.
- The underlying model matters less than getting the four components right. The same second brain has been run successfully on different models and harnesses with little change in quality.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
Why does a second brain need a framework at all?
The temptation when building one of these systems is to keep adding tools, integrations, and automations until it becomes unmanageable. That’s the most common failure mode: overcomplicating a system that’s supposed to make your life simpler. A framework forces a clearer question: what does this system actually need to do its job?
The job is narrow. A second brain should function as an extension of you, with enough knowledge of your business and enough reach into your tools that it can act like a capable teammate rather than a generic assistant. Everything that supports that goal matters. Everything else is noise. The four-component framework (context, connections, cadence, capabilities) exists to keep builders honest about that distinction, whether they’re setting up a system for the first time or auditing one that’s grown messy over time.
Importantly, this framework is tool-agnostic. It applies whether the system runs on Claude Code, Codex, OpenClaude, or any other harness. The components are your responsibility to build and maintain regardless of which model or platform sits underneath.
How does the “context” component work?
Context is the foundation, and arguably the most important of the four. It answers a simple question: does your second brain know your business without having to search for it?
A workable approach uses two layers. The first is a small set of core documents loaded into every single session. These typically include something like a “soul” file describing how the agent should behave, a “user” file documenting you and your business, and a “memory” file holding the handful of facts you want surfaced immediately, like current priorities, without a search. This top layer is what makes the system feel instantly familiar with your work.
The second layer is a deeper knowledge base, structured as an interconnected set of markdown documents (a wiki-style setup, sometimes referred to as a Karpathy-style LLM wiki). Every conversation with the agent gets summarized and logged, feeding into daily notes and entity documents for every project, person, or company involved in the business. This is the system’s long-term memory, searchable but not force-fed into every session.
A technical mechanism, like a session-start hook, pulls the top-level files into every new conversation automatically, so the agent always starts with the right baseline before it even looks at the deeper knowledge base.
The audit: open a brand new session and ask what your business does and who’s involved. If the agent has to dig or comes back vague, the context layer needs work. The fix is often as simple as telling the agent directly that it isn’t passing the test and working through improvements together. Iterating this way, in small daily conversations about what’s working and what isn’t, is how a second brain’s usefulness compounds over time.
How does the “connections” component work?
Context without reach is just a smart chatbot. Connections are what let a second brain act where the actual work happens: email, task management, file storage, code repositories, and team communication.
One coffee. One working app.
You bring the idea. Remy manages the project.
In practice, most platforms have an existing API or MCP server that can be wired in without much custom engineering. For platforms without an easy integration, a simple custom CLI tool can fill the gap. The goal is coverage across every place you actually operate, not a handful of convenient ones.
The audit: ask for a Monday morning catch-up, covering priorities, outstanding replies, and people who need attention. If the agent can’t assemble that quickly across your real tools, the connections layer isn’t complete yet.
One practical challenge with connections at scale is tool overload. Loading every available tool into an agent’s context up front can burn an enormous number of tokens before a single message is even processed. Dynamic tool recommendation, where an agent describes the task and receives only the relevant tools for that moment rather than the entire catalog, is one way to manage this as the number of connected platforms grows.
What do “cadence” and “capabilities” add to the system?
The first two components, context and connections, are about what the agent knows and where it can reach. Cadence and capabilities round out the picture: cadence covers the rhythm of how and when you interact with the system (daily check-ins, recurring summaries, standing routines), while capabilities cover the actual skills and outputs the agent can produce, like drafting documents, generating diagrams, preparing presentations, or researching a topic end to end.
A notable point from builders running these systems day to day: a modern coding agent can often handle capabilities like video editing, diagram creation, or presentation building directly through commands on your own machine, without needing a separate paid service for each function. That keeps the overall stack small even as the range of tasks the agent can handle keeps growing.
Is it worth building an AI second brain?
For anyone doing varied knowledge work, the case is straightforward: a well-built second brain can save a significant number of hours weekly by handling research, drafting, task triage, and routine production work across a business. The barrier to entry has dropped substantially because current coding agents and LLMs are capable enough that you don’t need a large, custom-built toolchain. A handful of markdown files, a coding agent, and a few connected platforms can go a long way.
The caveat is that it requires ongoing attention. These systems don’t become excellent on day one. They improve through repeated small audits against the four components and conversations where you explicitly tell the agent what isn’t working. The payoff scales with that invested iteration, not with the sophistication of the underlying model.
Frequently Asked Questions
What are the four components of an AI second brain?
Context, connections, cadence, and capabilities. Context is what the agent knows about your business by default, connections are the platforms it can act in, cadence is the rhythm of interaction, and capabilities are the actual tasks it can execute.
Do I need Claude Code specifically to build one?
No. The four-component framework applies regardless of the underlying harness or model, whether that’s Claude Code, Codex, or another agent platform. What matters is how well context and connections are built, not which LLM runs underneath.
What tools are commonly used for the knowledge base?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
A folder of markdown files is a common, simple approach, often managed in a note-taking app like Obsidian (Notion is another option). This keeps the knowledge base portable and easy to search without needing a dedicated database.
How do I know if my second brain’s context is good enough?
Open a brand new session and ask it what your business does and who’s involved. It should answer immediately and accurately, like a teammate would, without needing to search deeply through files or integrations first.
How do I know if the connections are set up correctly?
Ask for a summary of your priorities and outstanding items, similar to a Monday morning catch-up. If it can pull accurate, current information from your real tools (email, task manager, calendar) quickly, the connections are working as intended.
