AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What A Shared Second Brain Looks Like For AI Agents on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Thorsten Meyer describes a personal knowledge base that lets his AI agents share long-term memory through Markdown files and a Git-tracked index. As of Nov. 21, 2026, he reports 1,278 pages and a 97.4% top-1 recall score; the setup is designed for one owner, and its performance has not been independently verified.

Thorsten Meyer has described a shared long-term memory for the AI agents he runs, built from Markdown files and a Git-tracked index rather than a vector database or cloud service. In an article published Nov. 21, 2026, Meyer reported that the system held 1,278 pages and had a 97.4% top-1 recall score, a measure he says tracks how often its first selected page is the right one.

Meyer says he built the system because agents typically lose access to session context when a conversation ends. Instead of re-explaining project details, decisions and failed approaches in each new session, an agent can consult the same knowledge base as the others. Agents can also add information to it. The files are organized by clusters of related facts, such as a project or decision, and can be opened in a standard text editor.

Retrieval follows an index-first process: a script called recall.py scores the index entries without opening every memory file, then opens the best-matching page and returns the section it identifies as relevant, along with its source. Meyer says that source reference lets him inspect the underlying page. A separate script, remember.py, writes a new page, its index entry and a log entry together.

The article lists 1,761 links among the 1,278 pages and 1,363 recalls as of Nov. 21. Meyer also describes nine system views, including a map of connections, a daily brief, a triage process for session notes, a change log and a live view of agents recalling and adding memories. These are features of his own setup, not reported results from a comparative test.

At a glance
reportWhen: Published Nov. 21, 2026; system figures…
The developmentThorsten Meyer published a description of a shared, file-based memory system for the AI agents he runs.

One Memory Across Separate Agents

A shared store offers a practical alternative to giving each agent a separate, hard-to-inspect memory. If the design works as Meyer describes, a decision recorded during one session can be available to another agent without the user repeating it. Keeping the material in readable files also gives the owner direct access to revise or remove a memory and inspect the history of changes.

The approach trades managed infrastructure for owner control and portability. Text files and Git can be inspected with familiar tools, but the system depends on accurate indexing and consistent workflows. Meyer’s reported recall score is specific to his system and measurement; the article does not establish that similar results would follow in other projects or with other agents.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

From Session Notes to Lasting Memory

Meyer frames the problem around the limits of an agent’s active context: information available during a session may not be available when that session ends. He says some existing memory features are difficult for him to read, edit or audit, and may be separate across tools. His stated goal was one memory that he could inspect himself and share among the agents he operates.

In his workflow, material left behind after sessions is not automatically treated as permanent knowledge. A triage view sorts captures into items to preserve, merge with existing pages or discard. The article also describes an almanac for reviewing page use and system health, and a chronicle that records changes. Those elements show that the setup is not only a folder of notes; it also includes processes for maintaining them.

“I got tired of being their memory. So I built them one.”

— Thorsten Meyer, writing on ThorstenMeyerAI.com

Limits of the Reported Results

The figures and performance measure come from Meyer’s account. The article does not provide an independent evaluation, a detailed test protocol for the 97.4% top-1 score, or a comparison with other memory systems. It is therefore unclear how the score was calculated, what kinds of questions were tested, and how well the system performs on tasks beyond the author’s own use.

The description also does not set out a team-access model, permissions for multiple users, or safeguards for conflicts when several agents update related information. Meyer explicitly characterizes the setup as designed for one owner. The supplied article text ends during its discussion of the system’s statistics, so any additional figures or qualifications beyond those listed above cannot be confirmed here.

Testing the Design Beyond One Owner

Meyer’s article points readers to an explainer video showing the system’s views and workflow, but it does not announce a release schedule, public repository or planned product launch. The next steps for readers considering a similar setup would be to inspect the files and scripts, define how memories are reviewed and corrected, and test retrieval against their own questions rather than assume the reported score will transfer.

For the system described, the practical test is whether agents can consistently find accurate, current information while preserving a traceable source. The article does not state that Meyer plans to adapt it for teams or publish further benchmark results.

Key Questions

What is a shared second brain for AI agents?

In Meyer’s description, it is a long-term knowledge base that both he and his agents can read and update, so information can carry across separate sessions.

How is Meyer’s system stored?

It uses Markdown files kept in Git, with one index file listing each memory. Meyer says the setup does not use a vector database or cloud service.

How does an agent retrieve a memory?

A script scores entries in the index, opens the best-matching page and returns a relevant section with its source, according to Meyer. It does not search every page’s full contents first.

How accurate is the reported recall score?

Meyer reports a 97.4% top-1 score as of Nov. 21, 2026, describing it as how often the first page selected is the right one. The article does not provide an independent test or enough methodology to assess how broadly that figure applies.

Is the system ready for teams?

That is not established. Meyer says the setup is built for one owner, not a team, and the article does not describe shared permissions or multi-user conflict handling.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

Following The Food Search Trend Around Mountain Mike’s Pizza

An IdeaNavigator AI brief flags a Google Trends signal for Mountain Mike’s Pizza and proposes testing a food-industry monitoring service.

The Hidden 80%: AI’s Service Turn

A report traces the growth of U.S. service work and Baumol’s cost disease, then examines evidence that AI can raise customer-support productivity.

The Three Settings That Help Explain Different Jobs

A new ThorstenMeyerAI.com series opens with the three-sector model and data on how employment has moved from farming to services.

Supplier Or Shareholder? TSMC, Nvidia And Google Take Both Roles

TSMC approved up to $500 million each for Anthropic and OpenAI, while seven companies agreed to buy $26 billion of Anthropic shares after its IPO.