🔍 Read the full analysis: AI Agents And The Second Brain: A Shared Memory Approach on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Thorsten Meyer says he built a shared “Second Brain” so the AI agents he runs can draw on the same long-term memory across sessions. The system uses Markdown files, a Git history and an index; Meyer reports 1,278 pages and a 97.4% top-page recall score as of November 21, 2026. The figures come from his own account, and he says the setup is designed for one owner rather than a team.
Thorsten Meyer has described a shared “Second Brain” that lets the AI agents he runs read and add to one long-term memory, rather than starting each session without knowledge of earlier work. In an article published November 21, 2026, he said the system contained 1,278 pages and had a 97.4% top-1 recall score, figures that he reports for his own setup rather than independent testing.
The system stores knowledge in Markdown files kept in Git, with each page covering a cluster of related facts, such as a project or decision. A single INDEX.md file lists each memory on one line. Meyer says the setup uses neither a vector database nor a cloud service, and that the files can be opened in a standard text editor. His account adds to what he has shared about AI memory in practice.
When an agent needs information, a script called recall.py scores the index lines first, then opens the best-matching page and returns the relevant section with its source. Meyer says this lets him check where an answer came from. A separate script, remember.py, writes a new page, its index entry and a log entry together, a process intended to keep the catalogue and stored notes aligned.
Meyer reports that, as of November 21, the system had 1,278 pages, 1,761 links and 1,363 recalls. He gives a 97.4% score for “top-1” recall, which he defines as how often the first page selected by the system is the right one. The article does not describe an external evaluation or provide further details about how the score was measured.
The approach addresses a practical limitation of session-based AI agents: information available in one session may not carry into the next. A shared store can let multiple agents reuse project details and prior decisions, reducing repeated setup and giving a user one place to correct errors. Meyer’s account focuses on continuity across agents and the ability to inspect the source behind recalled information.
Its use of ordinary files and version control also makes the memory more legible than a store that a user cannot readily inspect. Git can preserve a history of edits, while citations to individual pages give a route to verify an answer. These features may help with oversight, but they do not by themselves establish that the contents are accurate or that an agent will select the right memory in every case.
Meyer also identifies a key limit: the design is built for one owner, not a team. A shared folder and index do not automatically resolve access controls, conflicting edits, privacy rules or responsibility for maintaining information when several people contribute. The article presents an individual workflow, not evidence of a tested enterprise system.
Products Worth Considering

Logitech MX Brio Ultra HD 4K Webcam

Dell XPS Tower 8960 Desktop, Intel Core i7-13700, NVIDIA GeForce RTX 4060, 16GB DDR5, 512GB SSD

LG 34-inch UltraWide WQHD Curved Monitor, 120Hz, HDR10
As an affiliate, we earn on qualifying purchases.
Meyer says he runs multiple agents and grew tired of repeating the same context after sessions ended. In his account, an agent operates on what it can see in its current context window; once that session ends, the information is no longer available to the agent unless it is stored elsewhere. He says built-in memory features can be difficult to read or audit and may be specific to individual tools.
His alternative is a persistent folder of text files with a compact index and scripts for retrieval and storage. The article also describes nine views intended to show activity and organization, including a map of links, a daily summary, a review process for captured information, a change log and a live view of agents recalling and writing. Meyer says the views help track how knowledge moves from session captures into lasting pages.
The figures in the article are a snapshot dated November 21, 2026. They describe Meyer’s system and should not be read as a general benchmark for AI memory tools or a performance guarantee for other users.
““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
How Recall Performance Was Measured
The 97.4% top-1 score is reported by Meyer, but the article does not specify the test set, number of questions, evaluation method or who judged whether a selected page was correct. The reported 1,363 recalls are a raw count; the article does not give a time window or comparison baseline for that number. The figures therefore describe Meyer’s own system and cannot establish how it would perform under other conditions.
It is also unclear how the system handles sensitive information, permissions, simultaneous edits or mistakes that enter the memory. The article explains the normal write path but does not detail recovery procedures, security controls or independent review. Its description of a one-owner setup leaves open what changes would be needed for a team deployment.
The Workflow Beyond One Owner
Meyer’s article describes the system as it stood on November 21, 2026, and outlines views for reviewing new captures, tracking changes and monitoring agents. It does not announce a product launch, public release or independent evaluation, and it gives no schedule for further development. The next useful evidence would include fuller details on how the recall score is tested and how the system handles access, privacy and collaborative editing.
For readers considering a similar arrangement, the account offers a description of one working design, not a validated blueprint. Its stated trade-off is clear: text files and a single index are inspectable and portable, while keeping them accurate requires ongoing discipline. Whether the approach can be adapted safely and reliably for larger groups remains unreported.
Key Questions
What is Meyer’s Second Brain?
It is a shared long-term knowledge base that Meyer says his AI agents can read and update. Its memories are stored as Markdown files, with an index that points to each page.
How does an agent find a relevant memory?
According to Meyer, recall.py scores the index entries first, opens the best-matching page and returns the relevant section with a source reference. The article says the script does not search by opening every page.
What does the 97.4% figure mean?
Meyer calls it the top-1 recall score: how often the first page selected is the right one, by his system’s measure. The article does not explain the test method or report an independent assessment.
Is the system designed for teams?
No. Meyer says it is built for one owner, not a team. The article does not describe team permissions, shared editing procedures or other controls that a group deployment may need.
Source: ThorstenMeyerAI.com
