TL;DR
Thorsten Meyer says he built a shared “Second Brain” so his AI agents can draw on the same long-term memory instead of starting each session without his work history. As of Nov. 21, 2026, he reports 1,278 markdown pages and a 97.4% top-one recall score; the system is his own, and its performance has not been independently assessed in the source material.
The system stores information in plain markdown files kept in git, 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 setup uses no vector database or cloud service, and that he can open and edit the files with ordinary tools, as described in his account of memory research on AI agents.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
For a question, a small script called recall.py scores the index entries before opening a file. It selects the best-matching page and returns the relevant section with its source, according to Meyer. To add information, remember.py writes the page, its index entry and a log entry together, a design intended to keep the catalogue aligned with the stored pages.
Meyer reports 1,761 links among the pages and 1,363 recalls, alongside the 1,278-page total. He defines the 97.4% figure as the system’s own top-one measure: how often the first page selected is the right one. The article does not describe an independent evaluation or provide the test set and scoring method.
Why I Gave My AI Agents One Shared Memory
Thorsten Meyer’s “Second Brain” lets his AI agents read and update one shared store of work history. It aims to reduce repeated setup while keeping memories readable and inspectable.
One record for every agent
The personal knowledge base stores related facts in plain markdown files kept in git.
Agents use the same memory
Information added in one workflow can be available to another agent session.
Files you can inspect
Meyer says he can open and edit the records with ordinary tools.
How a memory is found
The system searches an index first, then opens the selected page.
recall.py ranks the one-line descriptions in INDEX.md.
It selects a page and returns a relevant section with its source.
remember.py writes the page, index entry and log entry together.
What the score tells readers
The 97.4% figure is Meyer’s own measure of how often the first selected page is right.
- Top-one recall score
- Page and recall totals dated November 21, 2026
- Independent evaluation
- Test set or scoring method
- Time saved or errors reduced
A personal workflow with upkeep
The index needs clear descriptions, and Meyer says the setup is built for one owner.
| Potential value | Open question |
|---|---|
| Shared project history across agents | Team permissions and shared editing |
| Readable records and edit history in git | Stale facts, sensitive information and recovery from mistakes |
A common memory can reduce the repeated setup Meyer describes: explaining the same project history, prior decisions and failed approaches each time an agent session begins. If several agents can consult the same material, information added during one workflow may be available in another. That is the practical case Meyer makes for treating memory as a shared resource rather than a feature tied to a single assistant.
The design also makes the records readable and inspectable, and git preserves a history of edits. Meyer says recalled answers include their source page, allowing him to check the material behind a response and correct a fact at its origin. These characteristics could matter to people who want more control over what agents retain, though they do not by themselves establish that every retrieved fact is accurate or current.
The trade-off is upkeep. The index depends on clear, useful one-line descriptions, and the system’s normal write path must be used to keep pages, index entries and logs together. Meyer says the setup is built for one owner, not a team; the article does not explain how access, conflicting edits or shared ownership would work at organizational scale.
Products Worth Considering

Amazon Kindle Scribe (16GB) with Premium Pen, Tungsten

Apple AirPods Pro (1st Generation) with MagSafe Charging Case

Jackery HomePower 3000 Portable Power Station, 3072Wh, 3600W
As an affiliate, we earn on qualifying purchases.
From Session Context to Files
An AI agent generally works with information available in its active session. Meyer’s post says that when a session ends, that context is gone, while memory features offered by different tools may be difficult to inspect or may remain separate. He presents the Second Brain as his response to those limits: a store he can read directly and make available to the agents he runs.
The system’s workflow is organized around an index-first search. Instead of opening every note, the script ranks index lines, selects a page and extracts the section it considers relevant. The accompanying explainer describes nine views, including a map of connections among projects, skills, routines and apps. The source excerpt also begins describing a morning brief, but does not provide enough detail to characterize the remaining views.
The article and explainer were published by Thorsten Meyer AI on Nov. 21, 2026. The source identifies the explainer’s narration as an AI voice clone of Meyer and says its diagram is generic rather than a screenshot. The system description and usage figures come from Meyer’s own account of his project.
““I got tired of being their memory. So I built them one.””
— Thorsten Meyer, describing his motivation
How Recall Was Measured
The 97.4% top-one score is reported by Meyer as the system’s own measure. The source does not specify the number or type of questions tested, how a correct match was judged, or whether the results were independently reviewed. It is therefore unclear how the figure would compare with other memory or retrieval systems.
The account also does not quantify time saved, errors reduced or the effort needed to maintain the index. It describes a setup for one owner, but leaves open how it would handle team permissions, simultaneous edits, sensitive information, stale facts or recovery from mistakes. The reported page and recall totals are dated Nov. 21, 2026, and may change as the system is used.
Further Detail on Use
Meyer’s post presents the system and points readers to an explainer video, but the source material gives no announced release, independent benchmark or deployment plan for other users. The next useful evidence would be a fuller account of the recall evaluation, examples of how the system handles incorrect or outdated memories, and details on the nine views introduced in the video.
For now, the reported development is a personal workflow: Meyer says he uses a shared markdown knowledge base with his own agents. Whether the approach works as well for other users, or can support teams with different access and editing needs, remains to be established.
Key Questions
What is Thorsten Meyer’s Second Brain?
It is a personal knowledge base of markdown pages that Meyer says he shares with the AI agents he runs. The agents can consult existing pages and add information to the same store.
How does an agent find a memory?
According to Meyer, recall.py scores entries in INDEX.md, opens the best-matching page and returns the relevant section with its source. It does not search every markdown page directly.
What does the reported 97.4% score mean?
Meyer describes it as the system’s own top-one recall score: how often its first selected page is the right one. The source does not provide the test method or an independent evaluation.
Can teams use the system?
Meyer says the current setup is designed for one owner, not a team. The article does not describe team permissions, shared editing or other organizational features.
Source: Thorsten Meyer AI
