🔍 Read the full analysis: How My AI Agents Build On One Another’s Memory on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- 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 consult and add to the same long-term memory. As of November 21, 2026, it contained 1,278 pages; Meyer reports a 97.4% top-1 recall score, but the source does not provide an independent evaluation or full testing details.
Thorsten Meyer says he built a shared, long-term memory system for the AI agents he runs, allowing them to look up and add information across sessions, a topic explored in research on AI agent memory. In an article published November 21, 2026, he describes a Git-based collection of Markdown files and reports that it held 1,278 pages, with a 97.4% top-1 recall score on the system’s own measure.
Meyer calls the system his Second Brain, a way to manage agent memory like the approaches discussed in reducing memory costs. Each page holds a cluster of related facts, such as information about a project or decision. An INDEX.md file lists every memory in one line, and a recall script scores those lines before opening the best-matching page. The system then returns the relevant section and its source, according to Meyer.
To add information, a separate script called remember.py writes the memory page, its index entry and a log entry together. Meyer says this process is intended to keep the index aligned with the stored notes. Git records changes over time; he says the system does not use a vector database or cloud service.
The article reports 1,278 pages, 1,761 links and 1,363 recalls as of November 21, 2026. Meyer gives a 97.4% top-1 score, describing it as how often the first page selected is the right one. The source does not specify the evaluation set, how “right” is judged or the period covered by the recall count.
The system addresses a practical limitation of agents that lose session context: people may otherwise need to repeat project details, decisions and failed approaches. A common, inspectable store could let different agents draw on the same information rather than keeping separate memories inside individual tools.
Meyer emphasizes that the material is readable and editable as ordinary files, and that recalled answers include a source page. That can make it easier for the owner to check or correct information. The trade-off he identifies is maintenance discipline: the index and pages must be kept useful, and the setup is designed for one owner rather than a team.
The reported performance figure is a result from Meyer’s system, not an independently verified benchmark. It offers a snapshot of his implementation, but readers cannot use it to compare accuracy with other memory products without more information about the test method and conditions.
Products Worth Considering

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

Jackery HomePower 3000 Portable Power Station, 3072Wh, 3600W

DJI Mini 4K Drone with 4K UHD Camera
As an affiliate, we earn on qualifying purchases.
Why Meyer Built a Second Brain
Meyer frames the project around the gap between an agent’s current context window and long-term continuity. When a session ends, information in that session may no longer be available to the agent. He says existing memory features can be difficult to inspect or edit and may vary from one tool to another; those are his reasons for building a shared alternative.
Beyond the underlying files and index, Meyer describes nine views for managing and observing the system. They include a map of connections, a morning brief, a triage process for session captures, a change log, a page explorer and live views of agents recalling or saving information. He also describes a daily control desk and a console for monitoring services and machines.
These views are part of Meyer’s own setup, rather than evidence that every feature is required to use the basic approach. The core method he outlines is narrower: store related facts in plain-text pages, maintain a one-line index and have agents search that index before retrieving a page.
“I got tired of being their memory. So I built them one.”
— Thorsten Meyer, in the November 21, 2026 article
How the Recall Score Was Tested
The article does not explain the test set, scoring procedure or evaluation period behind the 97.4% top-1 figure. It also does not say how the 1,363 recalls were counted or whether they cover a defined time window. The figures should be treated as Meyer’s reported system metrics, not an independently audited result.
Other operational details are not fully specified, including how conflicts between memories are handled, how information is protected beyond the statements about personal data in the video, and how the system would manage permissions or shared edits in a team. Meyer says the design is intended for one owner, so its suitability for collaborative use remains unestablished.
The System’s Next Test Is Use
Meyer’s article explains the design and reports a snapshot of its size and recall performance, but it does not announce a product launch, public release or independent evaluation. The source also gives no timetable for future testing or development.
For readers considering a similar setup, the next practical step would be to assess whether the index, page structure and update process remain accurate as the collection grows. Any comparison of performance would require a defined test set, a clear scoring method and a stated measurement window. Meyer’s report does not say whether he plans to publish those details.
Key Questions
What is Meyer’s Second Brain?
It is a shared collection of long-term memories for the AI agents Meyer runs, stored in Markdown files with a central index, according to his article.
How do the agents find a memory?
Meyer says a script scores the lines in INDEX.md, opens the best-matching page and returns the relevant section with its source.
What does the 97.4% figure mean?
Meyer describes it as the system’s top-1 recall score—how often the first page selected is the right one by his measure. The article does not give the test method or an independent verification.
Does the system use a vector database or cloud service?
Meyer says it uses plain Markdown files kept in Git, with no vector database or cloud service in the described setup.
Is it designed for teams?
Meyer says the setup is built for one owner, not a team. The article does not establish how it would handle team permissions, conflicting edits or shared governance.
Source: ThorstenMeyerAI.com
