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

TL;DR

Thorsten Meyer published an account of a shared memory system that lets his AI agents read and add to the same collection of Markdown files. He says it held 1,278 pages and scored 97.4% on his top-one recall measure as of Nov. 21, 2026; the figures come from his own system, and the setup is designed around one owner.

Thorsten Meyer published a description on November 21, 2026, of a shared “Second Brain” that lets the AI agents he runs consult and update one collection of Markdown files. Meyer says the system contained 1,278 pages and had a 97.4% top-one recall score on that date, offering a detailed example of how one operator handles continuity between agent sessions.

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 stores memories as plain text pages kept in Git, with each page covering a cluster of related information, such as a project or decision. A file called INDEX.md lists the pages, one line per memory. Meyer says the setup uses no vector database or cloud service, and that the files can be opened in a standard editor.

When an agent receives a question, a small script called recall.py scores the index lines first. It then opens the best-matching page and returns the relevant section with its source. Meyer says this lets him inspect the page behind an answer and correct a fact at its source. A separate script, remember.py, writes a new page, its index entry and a log line together.

Meyer reported 1,761 links and 1,363 recalls alongside the page count. He defines top-one recall as how often the first page selected is the right one. The source material does not give the test set, evaluation period or a comparison against other memory systems, so the percentage describes his own measure rather than an independently verified benchmark.

At a glance
reportWhen: Published November 21, 2026
The developmentThorsten Meyer published a description of a shared, file-based memory system for the AI agents he runs.
AI Memory · Briefing

How AI Agents Share a Second Brain

Thorsten Meyer describes a shared collection of Markdown files that helps his agents carry information between sessions. The setup keeps memories readable and editable by one owner.

97.4%
reported top-one recall
1,278
memory pages
1,761
links reported
1,363
recalls reported

How agents find a memory

The workflow connects plain text notes to answers agents can trace back to a source.

1
Keep notes in Markdown

Pages group related information, such as a project or decision.

2
List pages in INDEX.md

One line per memory helps the recall script search the collection.

3
Retrieve with recall.py

It scores index lines, then returns a relevant section and its source.

4
Add notes with remember.py

The script writes a page, index entry and log line together.

What the design offers

Readable

Ordinary files

The notes can be opened and edited in a standard editor, with changes tracked in Git.

Shared

Continuity between sessions

The agents Meyer runs consult the same memory, reducing the need to repeat background.

Traceable

Sources behind answers

Meyer can inspect and correct the page used for an answer.

Who the setup fits

For one owner
  • Read and edit memory pages directly.
  • Track changes with Git.
  • Review the source behind a recalled answer.
For teams
  • The described arrangement is designed around one owner.
  • Team permissions and shared editing are not detailed.
  • Conflict handling is not described.
Evidence caveatThe 97.4% score is Meyer’s own measure; the source gives no test set, evaluation period or independent comparison.
The takeaway
A shared, source-linked file collection can give one operator’s AI agents continuity while keeping memory inspectable and editable.

Shared Memory Across Agent Sessions

The design addresses a practical problem for people who use agents across repeated tasks: each new session may lack details from earlier work, including prior decisions, project names and approaches that failed. A shared store can give multiple agents access to the same references, while source-linked answers give the operator a way to check what the system used.

Keeping the information in ordinary files also makes the memory easier for its owner to read, edit and track over time. Meyer presents that control as a reason for choosing the format. The tradeoff is ongoing maintenance and indexing discipline: the index must describe pages well enough for the recall script to find them, and the published setup is built for one owner rather than a team.

Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024 Model) - Metallic Black

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

As an affiliate, we earn on qualifying purchases.

From Session Context to Stored Notes

An AI agent can use information available in its current context window, but Meyer says he found that information did not carry over reliably between sessions. He describes his motivation as not wanting to repeat the same background each time he started an agent. His response was to maintain one memory store that every agent he runs could consult.

The article outlines a four-part workflow: keep facts in Markdown pages, list them in one index, search the index before opening a page, and add new material through a script that updates the page, index and log together. Meyer says Git keeps the change history. His explainer also describes nine views of the system, including a map of connections among projects, skills, routines and the applications that read or write the memory. The available source excerpt does not detail all nine views.

““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-one score is not accompanied in the source material by details about the questions tested, the number of trials, the test period or independent review. It is not clear how performance changes when a question spans multiple pages or when relevant information is missing from the index.

The article also does not provide a comparative evaluation against built-in agent memory features or other retrieval methods. Meyer says the setup is designed for one owner; the source does not describe access controls, team workflows or how conflicting edits would be handled if several people maintained it.

Further Details on the Nine Views

Meyer’s explainer video is presented as a walkthrough of the system’s nine views, including its map and a morning brief. The source material does not state a scheduled release, independent evaluation or planned expansion to team use. Readers seeking more detail on those points would need the fuller walkthrough or a later update from Meyer.

Key Questions

What is Thorsten Meyer’s Second Brain?

It is a shared collection of Markdown memory pages that Meyer says the AI agents he runs can read and update.

How does an agent find a relevant memory?

According to Meyer, recall.py scores 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 defines it as the system’s top-one recall score: how often the first page selected is the right one. The article does not provide the test details or an independent comparison.

Can teams use the same setup?

Meyer says the current arrangement is built for one owner. The source does not explain team permissions, shared editing or conflict handling.

Source: Thorsten Meyer AI

You May Also Like

Small Business AI Automation Software: How To Save Big This Labor Day

Discover how affordable AI automation tools can cut costs and boost efficiency for small businesses this Labor Day, with practical tips on choosing the right solutions.

Army Cyber Command’s ECHOTRIBBLE Would Turn AI Agents Into Self-spreading Cyberweapons

Documents reviewed by The Intercept describe an Army Cyber Command project for autonomous agents that could move across networks and disable targets.

How We Detect Frontier Model Misalignment

OpenAI says expanded monitors now cover more frontier-model activity, with tests suggesting earlier alerts could have helped in the Hugging Face incident.

Instagram’s AI Makes Teen-Status Calls Globally. What If They’re Wrong?

Meta says Instagram will use AI globally to identify suspected teens and apply Teen Account limits. The company has not disclosed misclassification rates.