TL;DR
Thorsten Meyer says his AI agents use a shared knowledge base made of Markdown files stored in Git and a single index. As of November 21, 2026, he reports 1,278 pages and a 97.4% top-1 recall score; the method and its performance figures come from his own account.
Meyer’s system, which he calls his Second Brain, keeps information in plain text files, with each page covering a cluster of related facts, such as a project or decision. A single file, INDEX.md, lists each memory in one line. Meyer says the files are kept in Git, allowing him to review their history, and that the system uses no vector database or cloud service.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
When an agent needs information, a script called recall.py scores the index entries first. The system then opens the page it selects and returns the relevant section with its source, according to Meyer. To add a fact, another script, remember.py, writes the page, its index entry and a log entry together. Meyer says this process is intended to keep the index aligned with the stored pages.
Meyer reports that the collection had 1,761 links and had been recalled 1,363 times by November 21. He says its top-1 recall score—the rate at which the first page selected is the right one—was 97.4%. These are figures reported by the author about his own system; his article does not describe an independent evaluation.
Inside My AI Agents’ Shared Second Brain
Thorsten Meyer describes a personal knowledge base that lets his AI agents share stored project details and decisions. The figures and performance claims come from his own account.
How the shared memory works
Plain text files and scripts give agents a persistent reference across sessions.
recall.py scores entries in INDEX.md to find a relevant memory.
The system opens the selected Markdown page and returns a relevant section with its source.
remember.py writes the page, its index entry and a log entry together.
What the design offers
The file-based approach makes the stored knowledge visible and editable by its owner.
- Agents can consult shared project details and decisions.
- Git history lets the owner review changes.
- Ordinary Markdown files can be read and edited with common tools.
- The account describes a setup for one owner, not a team deployment.
- Team permissions, simultaneous edits and conflicting updates are not established.
- The article does not explain how outdated facts are handled.
How to read the performance claim
Meyer reports a 97.4% top-1 score, meaning the first selected page was right at that rate according to his account.
| What is reported | What is not detailed |
|---|---|
| Top-1 recall score: 97.4% | Evaluation method and test set |
| Collection size and recall count | Independent testing or comparison |
| Personal workflow design | Performance on other users’ questions |
The system addresses a practical limitation of agent workflows: information available during one session may not be available when another begins. A shared store lets agents consult the same project details and decisions, while source links give Meyer a way to check where a returned answer came from.
Keeping the knowledge in ordinary files also makes it possible for the owner to read and edit the records with common tools. That offers a degree of visibility and portability, while placing responsibility for maintaining clear pages and useful index entries on the person running the system. Meyer describes the setup as designed for one owner, and does not establish that it is ready for collaborative team use.
Products Worth Considering

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)

Amazon Kindle Scribe Colorsoft 64GB with Premium Pen

Jackery HomePower 3600 Plus Portable Power Station
As an affiliate, we earn on qualifying purchases.
From Sessions To Stored Notes
AI agents work with information available to them during a session. Meyer says that when a session ends, its context is gone, so he had to repeat details about earlier decisions and failed approaches when starting new work. He built the Second Brain to provide a persistent reference that his different agents could access.
In his November 21 article and accompanying explainer video, Meyer presents the system as a combination of an index, individual Markdown pages and scripts for retrieval and entry. He also describes nine views for examining the collection, including a map of connections among projects, skills, routines and the applications that use the memory. The supplied account does not detail all nine views.
“One memory, shared by every agent I run, that I can open and read myself.”
— Thorsten Meyer, describing the system in his November 21, 2026 article
Performance And Team Use
The reported 97.4% top-1 score is described as the system’s own measure. Meyer’s account does not specify the evaluation method, test set or time period behind that score, so readers cannot assess how it would compare with other memory systems or perform on different questions.
It is also unclear how the system handles conflicting or outdated facts, permissions, simultaneous edits, or information that should not be shared across agents. Meyer characterizes the setup as built for one owner; his article does not report a team deployment or independent testing.
Further Details On The System
Meyer’s explainer video presents the system’s retrieval and writing process and previews its views. The article does not announce a product release, public repository, or planned expansion. Further information about evaluation methods, maintenance and team access would be needed to judge how readily the approach could be adapted beyond his personal workflow.
Key Questions
What is Thorsten Meyer’s Second Brain?
It is a shared knowledge base that Meyer says his AI agents can read from and add to. It consists of Markdown files, an index and scripts for retrieval and entry.
How does an agent find a relevant memory?
According to Meyer, recall.py scores the lines in the index, opens the best-matching page and returns the relevant section with its source.
How large was the system in November 2026?
Meyer reported 1,278 pages and 1,761 links as of November 21, 2026. He also reported a 97.4% top-1 recall score, without describing an independent evaluation in the supplied article.
Is the system designed for teams?
Meyer describes it as designed for one owner. His account does not establish how it would handle team permissions, shared editing or conflicting updates.
Source: Thorsten Meyer AI
