What Is An AI Second Brain, And How Does It Work?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What Is An AI Second Brain, And How Does It Work? on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

  • Fast, free delivery for dorm and study essentials
  • Prime Video and Amazon Music included
  • Member-only deals
Try Prime for Young Adults Free trial for eligible 18–24 year olds
As an affiliate, we earn on qualifying purchases.

TL;DR

Thorsten Meyer describes a personal knowledge base that lets his AI agents share information across sessions. It uses Markdown files stored in Git, an index for retrieval, and scripts for adding and finding memories. Meyer reports 1,278 pages and a 97.4% top-1 recall score as of Nov. 21, 2026; those figures come from his own system, not an independent evaluation.

Thorsten Meyer has described a shared memory system for his AI agents, built from Markdown files and Git, to help them reuse information across sessions instead of starting with no knowledge of his work. In an article published Nov. 21, 2026, Meyer said the system contained 1,278 pages and had a 97.4% top-1 recall score by its own measure; these are figures he reports for his personal setup, not independently verified results.

Meyer calls the system his Second Brain. It is a personal knowledge base that he and the AI agents he runs can both read and update. The system stores information in ordinary Markdown files tracked in Git, with each page covering a cluster of related facts, such as a project or decision. Meyer says it does not use a vector database or cloud service.

Retrieval begins with INDEX.md, a catalogue with one line describing each memory. Meyer says a script called recall.py scores those lines, opens the best-matching page and returns the section it judges relevant, along with the source page. The design is intended to let a user inspect the material behind an agent’s answer rather than rely on an opaque memory store.

A separate script, remember.py, writes a new fact to its page, adds the corresponding index line and records the change in a log. Meyer says this single-step process is meant to prevent the index from falling out of sync with the files. He also describes nine interfaces for monitoring and reviewing the system, including views for recent activity, links between pages and captured information that may need to be retained or discarded.

As of November 21, 2026, Meyer reported 1,278 pages, 1,761 links and 1,363 recalls. He defined top-1 recall as how often the first page selected by the system is the right one and gave its score as 97.4%. The source does not describe an independent test, the evaluation set or how “right” was judged.

At a glance
reportWhen: Published November 21, 2026
The developmentThorsten Meyer published a description of his personal AI memory system, explaining its file-based design, retrieval process and reported usage metrics.

Making Agent Memory Inspectable

The system addresses a practical limitation of conversational AI: information available during one session may not be available to an agent in a later session. A shared store can reduce repeated explanations and allow different agents to draw on the same project details, preferences or prior decisions. That could matter to people using several agents or tools on related work.

Meyer’s approach puts readability and control ahead of a specialized hosted memory product. Plain files can be opened and edited with common software, while Git records changes over time. Returning a source page with a recalled answer also gives the user a way to check what the system relied on. These features may help with auditing and correcting errors, but they do not establish that the retrieved information is complete or accurate.

The trade-off is maintenance. The index must describe pages well enough for the retrieval script to find them, and the stored knowledge must be reviewed as circumstances change. Meyer also says the design is for one owner rather than a team. Shared use across people would raise additional questions about permissions, conflicting edits and private information.

LG OLED evo C4 55-Inch 4K Smart OLED TV (OLED55C4PUA) with 144Hz Refresh Rate, HDMI Cable, 24-Month Protection Plan, and Streaming Bundle

LG OLED evo C4 55-Inch 4K Smart OLED TV (OLED55C4PUA) with 144Hz Refresh Rate, HDMI Cable, 24-Month Protection Plan, and Streaming Bundle

Samsung 65-Inch 4K OLED Smart TV with Dolby Atmos, 120Hz Refresh Rate, and Amber Protection Bundle

Samsung 65-Inch 4K OLED Smart TV with Dolby Atmos, 120Hz Refresh Rate, and Amber Protection Bundle

As an affiliate, we earn on qualifying purchases.

From Session Context to Stored Files

An AI agent’s working context is the information available to it during a particular interaction. Meyer’s article describes the problem from his own use: when a session ends, its context is not automatically carried into the next one, so he found himself repeating details to his agents. Some products offer built-in memory, he writes, but he wanted a system he could read, edit and use across the agents he runs.

The solution he presents is not a general standard or a claim that all AI systems work this way. It is one practitioner’s implementation: text files hold the memories, an index narrows the search, and scripts handle recall and updates. The system’s Git history supplies a record of file changes, while its log and monitoring views provide additional ways to follow activity.

Meyer also describes a process for deciding what should become lasting knowledge. Agent sessions can leave captures for review; a triage view is used to decide whether information should be saved, merged into an existing page or dropped. That human-facing step recognizes that not everything said in a session belongs in long-term memory.

“One memory, shared by every agent I run, that I can open and read myself.”

— Thorsten Meyer, in his November 21, 2026 article

Limits of the Reported Score

The performance figures are Meyer’s own measurements. His article does not specify the test questions, the number or selection of evaluated recalls, the judging process, or whether results were checked by an outside party. The reported 97.4% score therefore describes his system according to his stated measure; it cannot by itself establish how the approach would perform for other users, data or agents.

The article also does not provide a detailed account of privacy safeguards, access controls or recovery procedures. Keeping files locally and avoiding a cloud service are design choices Meyer reports, but they do not alone settle how sensitive information is protected, especially if agents or machines that can access the files are compromised. The article gives no comparative evaluation against commercial memory features or other retrieval methods.

It remains unclear how easily the setup can be adapted for teams, how much human review it requires as the collection grows, and how it handles contradictory or outdated facts. Meyer’s description is a first-person account of a working system, not a controlled study or an independent audit.

Testing Beyond One User

Meyer’s article does not announce a release date, product launch or formal evaluation. It presents the system as an account of his existing workflow, with figures current to November 21, 2026. Readers interested in adopting a similar approach would need to decide how to structure pages, keep the index accurate and review captured material.

For the reported results to support broader conclusions, useful next steps would include publishing the recall test method, explaining how answers are judged and comparing performance across different tasks and memory collections. A team-oriented version would also need clear controls for editing, access and conflicting information. Until such details are available, the system is best understood as a personal implementation whose reported benefits and limitations have not been independently assessed.

Key Questions

What does “AI Second Brain” mean in Meyer’s article?

It refers to a long-term knowledge base that Meyer and his AI agents can read and update. It is intended to preserve useful information between sessions.

How does the system find a relevant memory?

Meyer says the recall script first scores one-line descriptions in INDEX.md, then opens the best-matching page and returns a relevant section with its source.

Does it require a vector database or cloud service?

According to Meyer, no. His setup uses Markdown files in Git and scripts for recall and writing. That describes his implementation, not every AI memory system.

Is the 97.4% recall score independently verified?

The article provides no independent evaluation or detailed test method. The 97.4% figure is Meyer’s reported top-1 score for his own system, and the basis for judging the selected page is not fully explained.

Can a team use the system as described?

Meyer says the setup is built for one owner, not a team. His article does not describe multi-user permissions or processes for resolving conflicting edits.

Source: ThorstenMeyerAI.com

You May Also Like

Why Desktop Automation Benefits From Watch-Once Voice Commands

New approach leverages watch-once speech recognition for desktop workflows, aiding power users with repeated tasks and enhancing productivity.

Never Lose Support: The Importance Of Monitoring AI Operations Signals

Ensuring AI support remains active is crucial for operations teams. A new focus on signal monitoring helps detect shifts early, preventing disruptions.

A New Food Trend Emerges: Taco Bell’s Ice Cream Taco On Google Charts

A new food trend, Taco Bell’s Ice Cream Taco, has surged on Google Trends, signaling emerging consumer interest and potential market shifts.

Proton’s AI Plans: What’s On The Record And What’s Still Unknown

Proton canceled planned AI-use re-interviews but says it will keep pushing adoption. A reported internal code push raises questions about security review.