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

TL;DR

Thorsten Meyer published a description of a shared memory system for the AI agents he runs. It stores information in Markdown files tracked with Git, uses an index to select a page, and reports a 97.4% top-1 recall score as of November 21, 2026. The figures and performance measure come from the author; independent verification details are not provided.

Thorsten Meyer published details on November 21, 2026 of a shared memory system for the AI agents he runs, using Markdown files tracked in Git and a single index to help agents retrieve information about his work. The author says the system contained 1,278 pages and scored 97.4% on its top-1 recall measure at publication, though the article does not provide independent validation of those figures.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

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

The system, which Meyer calls his Second Brain, is intended to address the loss of information between AI sessions. He says agents otherwise begin without knowledge of prior decisions, failed approaches or project terminology, requiring him to repeat context. His system makes the same collection available to all the agents he runs, and allows them to look up and add information.

Each memory is stored as a plain-text Markdown page, with one index file, INDEX.md, listing each page in a line. When an agent receives a question, a small script scores the index lines first, then opens the best-matching page and returns the relevant section with its source. Meyer says the system uses no vector database or cloud service, and that the files can be opened in an ordinary editor.

A second script, remember.py, is described as writing a new page, its index entry and a log entry together. This is meant to keep the index aligned with the stored notes. Meyer reports 1,761 links and 1,363 recalls as of November 21, 2026. He defines top-1 as how often the first page selected is the right one, and reports a score of 97.4%; the explainer does not specify the evaluation set or testing method.

At a glance
reportWhen: Published November 21, 2026
The developmentThorsten Meyer published an explainer on November 21, 2026, detailing a shared, file-based memory system for his AI agents.
AI Memory · Briefing

What Is a Second Brain?How AI Agents Share One Memory

Thorsten Meyer describes a readable, shared memory system for his AI agents. It helps carry project context across sessions, with reported results that have not been independently verified.

97.4%
Author-reported top-1 recall
1,278
Memory pages at publication
1,761
Links reported
1,363
Recalls reported

How agents find a memory

The system searches its catalogue first, then opens a matching note and returns the relevant section with its source.

1
Score the index

A script ranks the lines in INDEX.md against the question.

2
Open the best match

The agent reads the selected Markdown page.

3
Return sourced context

The answer includes the relevant section and its source.

Why a file-based memory

One collection is available across the agents Meyer runs and remains readable to its owner.

Readable

Plain-text notes

Memories live in Markdown files that can be opened and corrected in an ordinary editor.

Traceable

Changes in Git

Git tracks file history, so the owner can review changes to stored knowledge.

Consistent

Add notes together

remember.py writes a page, its index entry and a log entry together.

What the score can tell you

Top-1 means how often the first selected page is the right one, according to Meyer.

Reported
  • 97.4% top-1 recall
  • 1,278 pages at publication
  • Performance figures from the author
Not specified
  • Evaluation set or testing method
  • Independent assessment
  • How well it works for other users or tasks

Limits for wider use

The explainer presents a system built around one owner.

Open questionsTeam permissions, shared editing, conflicting or outdated memories, and multi-user scale are not described.
The takeaway
A searchable, Git-tracked memory can give one owner’s AI agents shared context across sessions, but the reported recall score needs a disclosed evaluation to judge its broader value.

A Portable Memory for AI Agents

The design makes agent memory more readable and inspectable than a store hidden inside an individual product: an owner can open the source file, review a change in Git and correct a fact directly. Sharing the same collection across agents may also reduce repeated setup when work moves between tools or sessions.

The trade-off, as Meyer describes it, is that this approach depends on maintaining clear index entries and is built around one owner, rather than team access controls or shared editing. The reported recall score offers a snapshot of the author’s own system, but without a disclosed benchmark it cannot establish how well the approach would work for other users or tasks.

Products Worth Considering

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black

DJI Mic 3 Wireless Microphone (2 Transmitters, 1 Receiver, Charging Case)

DJI Mic 3 Wireless Microphone (2 Transmitters, 1 Receiver, Charging Case)

As an affiliate, we earn on qualifying purchases.

From Session Context to Stored Notes

An AI agent can use information present in its current context window, but that context may not carry over when a session ends. Meyer says he wanted a single memory he could read himself, available across the agents he uses, with answers traceable to a source page.

The article presents a compact workflow: index first, files second. Git supplies a history of file changes, while the index is the retrieval catalogue. The explainer also refers to nine views for inspecting the system, including a map of projects, skills, routines and connected apps. The supplied source text cuts off during its description of a morning brief, so it does not detail all nine views.

“One memory, shared by every agent I run, that I can open and read myself.”

— Thorsten Meyer, describing his system

How Recall Performance Was Measured

The reported 97.4% top-1 score is attributed to the system’s own measure, but the source does not explain how many questions were tested, how they were selected, or who judged whether a selected page was correct. It is not an independently assessed benchmark.

The article also does not specify how agents are granted access, how conflicting or outdated memories are handled, or how the setup would scale to multiple people. Those details matter for organizations considering a shared memory beyond a single owner.

Further Detail on the Nine Views

Meyer’s explainer video is described as walking through nine views of the system, but the available article text names only the map and begins describing a morning brief. More information about the remaining views, the scripts and the recall evaluation would clarify how the system works in daily use. No independent evaluation or team deployment is announced in the source.

Key Questions

What does Meyer mean by a second brain?

He means a long-term knowledge base shared by the AI agents he runs, as well as readable by him. It stores information about his work for retrieval across sessions.

How does an agent find a memory?

According to Meyer, a script scores the lines in INDEX.md, opens the best-matching Markdown page and returns the relevant section with its source.

What does the 97.4% figure measure?

Meyer describes it as the system’s top-1 recall: how often the first page selected is the correct one. The source does not give the testing method or an independent assessment.

Can a team use the same setup?

The article characterizes the setup as built for one owner. It does not describe team permissions, collaborative editing or a multi-user deployment.

Source: Thorsten Meyer AI

You May Also Like

Why Your Local LLM Feels Dumber Than It Is

Exploring why local large language models often appear less capable, with confirmed technical reasons and implications for users and developers.

The Ultimate Guide To Automation-Flow Rebuilders In Email Migrations

A new automation-flow rebuilding tool aims to streamline email platform migrations for agencies, reducing manual effort and errors during complex workflows.

ChatGPT Work Tool And Skill Reference

OpenAI introduces a new ChatGPT Work Tool and Skill Reference to enhance professional use and skill development, sparking increased interest among users.

What The GenAI R&D Head’s Exit Means For ByteDance’s Embodied Intelligence Push

36Kr reports ByteDance’s GenAI R&D head has resigned for embodied intelligence, alongside three core-member departures in four months.