🔍 Read the full analysis: What It Takes To Build A Second Brain For AI Agents on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Thorsten Meyer describes a personal memory system shared by his AI agents, using Markdown files in Git and a single index rather than a vector database or cloud service. As of Nov. 21, 2026, he reports 1,278 pages and a 97.4% top-1 recall score, but the score is his system’s own measure, and the setup is designed for one owner rather than a team.
Thorsten Meyer has described a shared long-term memory system that lets his AI agents consult and add to the same collection of notes, rather than starting each session without personal work history. In an article published Nov. 21, 2026, Meyer said the system uses Markdown files stored in Git and a single index, and reported a 97.4% top-1 recall score on its own measure.
The system, which Meyer calls his Second Brain, stores related facts together in individual pages, such as information about a project or a decision. A file called INDEX.md lists each memory on one line. Meyer says the setup uses neither a vector database nor a cloud service, and that the text files can be opened and edited with ordinary software.
When an agent needs information, a script called recall.py scores the index entries first. It then opens the best-matching page and returns the relevant section with its source, according to Meyer. A separate script, remember.py, writes a new page, its index entry and a log entry together. Meyer says this is intended to keep the index and the stored material aligned.
As of Nov. 21, Meyer said the system contained 1,278 pages and 1,761 links, and had recorded 1,363 recalls. He reported a 97.4% top-1 score, which he defines as how often the first page selected by the system is the right one. The article does not describe an independent evaluation or provide a benchmark against other memory systems.
The design addresses a practical limitation of session-based AI agents: information visible in one session may not be available in the next. A shared store can let agents retrieve project decisions and prior work without requiring the user to repeat them, while source links give the user a way to inspect the material behind a response.
Keeping memories in plain, version-controlled files also makes them readable and editable outside a particular agent product. Git provides a history of changes, Meyer says. That can support review and correction, though it does not by itself guarantee that stored facts are accurate or that an agent will retrieve the right one.
The approach has a stated limit: Meyer describes a system built around one owner, not a shared team workflow. Organizations considering a similar design would still need to decide how multiple people can contribute, manage access, resolve conflicting edits and handle sensitive information.
Popular right now

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

KitchenAid 5.5-Quart Bowl-Lift Stand Mixer, Porcelain White

Vitamix VX1 Professional-Grade Blender, 64 oz, Black
As an affiliate, we earn on qualifying purchases.
From session context to stored memory
AI agents work with information available in their active context, and that context does not necessarily carry over after a session ends. Meyer says he wanted a memory that he could read and correct himself and that would be available to all the agents he runs, rather than separate, less accessible memory stores tied to individual tools.
His system also includes nine views intended to help him monitor and manage the collection. These include a map of connections, a daily brief, a triage process for deciding which session captures become durable memories, and a chronicle of changes. Other views show links between pages, activity as agents read and write, and the status of services running the system. These are features of Meyer’s implementation, not evidence that every agent-memory system needs the same components.
““One memory, shared by every agent I run, that I can open and read myself.””
— Thorsten Meyer, describing the system
How the recall score was measured
The reported 97.4% top-1 score is a measure defined by Meyer for his own system. His article does not provide the test set, number of test queries, scoring procedure or an external review, so readers cannot determine from the material how the result would compare with other systems or perform on different users’ data.
It is also unclear how the system handles incorrect or outdated memories beyond allowing changes to the files and recording them in a log. The description does not set out a team permissions model, privacy controls, backup arrangements or safeguards against an agent writing misleading information into the store. Those details would matter for broader organizational use.
Testing the design beyond one owner
Meyer’s article describes the system as it stood on Nov. 21, 2026, but does not announce a product launch, release schedule or independent evaluation. The next useful evidence would be more detail on how recall is tested, how often retrieved pages are correct in varied situations, and how the system behaves as the memory grows.
For readers adapting the idea, the immediate work is operational: define what agents may store, review new entries, correct stale information and decide how changes are tracked. A team version would also need clear ownership and access rules. Meyer presents the setup as a personal system, so its suitability for shared or regulated work remains unestablished.
Key Questions
What is Meyer’s Second Brain?
It is a personal knowledge base shared with his AI agents. Agents can look up information in it and, through a separate process, add new memories.
How does the system store and find information?
It stores related facts in Markdown pages in Git. A one-line-per-memory index is scored first; the recall script then opens the best-matching page and returns a relevant section with its source, according to Meyer.
What does the reported 97.4% score mean?
Meyer defines top-1 recall as how often the first page selected is the right one. The article reports 97.4% as of Nov. 21, 2026, but does not provide an independent test or enough methodology to compare the figure with other systems.
Can a company use the same setup?
The underlying files and index could be adapted, but Meyer says his version is built for one owner, not a team. Shared permissions, conflict handling, privacy and review procedures are not detailed in the article.
Source: ThorstenMeyerAI.com
