🔍 Read the full analysis: How To Give AI Agents A Second Brain They Can Share on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Thorsten Meyer says he built a shared long-term memory for his AI agents using Markdown files, Git and a one-line-per-page index. As of Nov. 21, 2026, he reports 1,278 pages and a 97.4% top-1 recall score; those figures come from his own system, and he describes it as a personal setup rather than a team product.
Thorsten Meyer has described a shared memory system for the AI agents he runs, built from plain Markdown files, Git and a searchable index. In an article published Nov. 21, 2026, he said the system holds 1,278 pages and has a 97.4% top-1 recall score, according to his own measurement; the approach aims to let agents carry useful information across sessions without relying on separate, hard-to-inspect memory stores.
Meyer calls the system his Second Brain. Each memory is stored as a Markdown page, generally grouping related facts about a project or decision. A single file, INDEX.md, lists the pages one line at a time. The files are kept in Git, which Meyer says preserves a history of changes. He says the setup uses no vector database or cloud service.
When an agent needs information, a script called recall.py scores the index entries first, then opens the best-matching page and returns the relevant section along with its source. Meyer says this lets him check which page informed an answer. Another script, remember.py, writes a new page, its index entry and a log entry together, a process intended to keep the index aligned with the stored notes.
Meyer reported 1,278 pages, 1,761 links and 1,363 recalls as of Nov. 21. He gives a 97.4% score for “top-1” recall, which he defines as how often the first page selected is the right one. The article does not provide the test method, evaluation set or measurement period behind that percentage, so it should be read as a result reported by the system’s creator, not an independently verified benchmark.
The design addresses a practical weakness of session-based AI work: when a session ends, its context may not carry into the next one. A shared store can give different agents access to the same project decisions, terminology and prior attempts, potentially reducing repeated explanations and inconsistent answers. Meyer presents this as a solution he built for his own daily use, not as a tested product or a claim that agents will always retrieve the correct information.
Keeping memories in readable files changes who can inspect and maintain them. A user can open a page, review its history in Git, and correct a mistaken entry directly. That readability and portability may matter to people concerned about opaque, tool-specific memory. But the approach also puts responsibility on users to write useful index lines, decide what should be retained and keep the memory organized.
Meyer says the setup is designed for one owner, not a team. Shared access across agents does not by itself settle questions about permissions, simultaneous edits, sensitive information or how to separate workspaces. Organizations considering a similar arrangement would need to address those issues rather than treating a personal folder as a ready-made team system.
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

Shark Pet Cordless Stick Vacuum, Up to 40-Minute Runtime, Grey, IX141

Dyson Cyclone V10 Origin Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
From Session Context to Stored Files
AI agents typically answer from information available in their active context window. Meyer’s account begins with the problem that details from a finished session may need to be explained again later. He says some agent tools offer memory, but he wanted one store he could read and edit and that could be used by the different agents in his workflow.
The system separates the memory into three practical parts: Markdown pages hold the information, an index helps agents locate a likely page, and scripts handle retrieval and new entries. Meyer also describes nine views for monitoring and managing the collection, including a map of connections, a daily brief, a triage process for session captures, a change chronicle and consoles for live activity and services. These are views of his implementation, not evidence that the underlying memory method has been independently evaluated.
The article’s main case is operational: keep agent memory in files that a person can inspect, and make retrieval return a source alongside the answer. The reported scale and accuracy figures describe Meyer’s own system at a specific date; the source material does not give a comparison with other memory approaches.
“One memory, shared by every agent I run, that I can open and read myself.”
— Thorsten Meyer, describing his goal
Limits of the Reported Recall Score
The 97.4% top-1 score is reported by Meyer, but the article does not explain how many test questions were used, how “the right page” was judged, or whether the score comes from a formal benchmark. It also does not compare performance with other retrieval systems or describe how the system handles questions for which no page is relevant.
The source does not establish how well the approach works for other users, larger organizations or agents operating concurrently. Details about access controls, backups, privacy protections and conflict handling are also not specified. Meyer says personal data shown in the accompanying video and images is protected, but that statement does not amount to a full account of the system’s security practices.
Adoption Depends on Workflow Discipline
Meyer’s article describes the system as it stood on Nov. 21, 2026; it does not announce a product launch, release schedule or independent evaluation. Readers seeking to use the approach would need to choose what information belongs in the memory, create clear page descriptions and establish a process for reviewing captured notes before they become lasting records.
The next meaningful evidence would be detail about the recall test, performance over time and how the design handles multiple users or conflicting edits. Until then, the reported figures offer a snapshot of one person’s implementation, while its broader value for teams remains an open question.
Key Questions
What is the Second Brain?
It is Meyer’s shared knowledge base for his AI agents: Markdown pages stored in Git, with an index that helps agents find information across sessions.
How does an agent find a memory?
Meyer says recall.py scores entries in the index, opens the best-matching page and returns the section it considers relevant, along with the source page.
Is the 97.4% recall figure independently verified?
The source article reports the figure as Meyer’s system measure. It does not describe the test method or provide an independent evaluation.
Can a team use the same setup?
Meyer describes the arrangement as built for one owner, not a team. The article does not explain team permissions, concurrent editing or other multi-user safeguards.
Source: ThorstenMeyerAI.com
