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

TL;DR

Thorsten Meyer described a personal system that lets his AI agents share a long-term memory stored in markdown files and tracked in Git. As of Nov. 21, 2026, he reported 1,278 pages and a 97.4% top-1 recall score; those figures come from the system’s creator, and no independent evaluation is provided.

Thorsten Meyer published a description on Nov. 21 of a shared memory system for the AI agents he uses, built from markdown files stored in Git. He says the system contains 1,278 pages and has a 97.4% top-1 recall score, offering one approach to preserving information across agent sessions while keeping the underlying records readable and checkable by their owner.

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.

Meyer calls the system his “Second Brain.” It is a personal knowledge base that he and his agents can both read and update. He says its records are plain markdown files, with each page holding a cluster of related facts, such as material about a project or a decision. A single file, INDEX.md, lists each memory in one line. Meyer says the setup uses neither a vector database nor a cloud service.

For recall, a script named recall.py scores the index lines against an agent’s question without opening all the stored pages. It then opens the best-scoring page and returns the section it identifies as answering the question, along with the source. Meyer says this lets him inspect where an answer came from. The process is described as deterministic, but the article does not provide technical details about how relevance is scored or how the chosen section is selected.

A second script, remember.py, writes a new page, its index line and a log entry together. Meyer says this single operation is intended to keep the index aligned with the stored pages. He reports that, as of Nov. 21, the system had 1,278 pages, 1,761 links and 1,363 recalls. He describes its 97.4% top-1 score as the rate at which its first selected page is the correct one, but does not give a test set, evaluation period or independent verification.

At a glance
reportWhen: Published Nov. 21, 2026
The developmentThorsten Meyer published a description of a personal markdown-based memory system shared by the AI agents he uses.
AI Memory · Briefing

What Happens When AI Agents Share a Memory?

Thorsten Meyer describes a personal memory system that lets his AI agents read and update the same markdown files. Its reported performance comes from the creator, with no independent evaluation provided.

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

One readable memory store

Meyer says he and his agents share a personal knowledge base tracked in Git.

Readable

Markdown pages

Related facts are grouped in plain text files that the owner can inspect and correct.

Indexed

A line per memory

INDEX.md lists each memory so agents can search descriptions before opening pages.

Tracked

Changes in Git

The file history lets the owner review how stored information changes over time.

How recall works

The described process searches the index first, then returns a matching section with its source.

1
Score the index

recall.py scores index lines against the agent’s question.

2
Open a page

The script opens the best-scoring page and selects a section.

3
Return answer and source

The agent gets the identified section and where it came from.

What this could mean

A shared record may reduce repeated explanations across sessions while keeping information accessible to its owner.

Potential value
  • Agents can draw on the same stored decisions and project terms.
  • People can read, correct, or move the plain files.
Open questions
  • Useful recall depends on organized pages and clear index descriptions.
  • The account does not establish team permissions or handling of conflicting edits.

Read the score with care

The article does not explain the test set, evaluation period, or what counted as the correct page.

CaveatThe 97.4% figure is Meyer’s own system measure, with no independent verification reported.
The takeaway
Shared, readable memory could help agents carry useful context across sessions, but the reported recall score cannot show how well the approach works elsewhere.

A Checkable Memory Across Agents

The system addresses a practical limitation of session-based AI work: information available in one session may not carry over when that session ends. Meyer says he was repeatedly explaining prior decisions, failed approaches and project terminology to agents, and built a shared store to reduce that repetition. If the approach works as described, one maintained record can serve several agents rather than requiring each tool or session to build a separate account of the same work.

Its use of plain files and Git also makes the records accessible to a person with a text editor and preserves a change history. That can help an owner inspect, correct or move the information. The tradeoff is upkeep: the index needs useful descriptions, and pages need to be organized well enough for the recall script to find relevant material. Meyer characterizes the setup as designed for one owner, not a team, so the account does not establish how shared editing, permissions or conflicting changes would work in a workplace.

The reported 97.4% score is a result from Meyer’s own system, not a general measure of AI-agent memory. Without details of the questions tested and what counted as a correct page, readers cannot judge how it would perform on other collections or tasks.

Products Worth Considering

Logitech MX Mechanical Mini for Mac Wireless Keyboard, Space Gray

Logitech MX Mechanical Mini for Mac Wireless Keyboard, Space Gray

Bose QuietComfort Ultra Wireless Noise-Cancelling Over-Ear Headphones with Spatial Audio, White Smoke

Bose QuietComfort Ultra Wireless Noise-Cancelling Over-Ear Headphones with Spatial Audio, White Smoke

As an affiliate, we earn on qualifying purchases.

From Session Context to Stored Notes

An agent’s context window holds information available during its current session. Meyer says that once a session ends, the agent may no longer have that information, prompting him to restate facts in later sessions. He also notes that some tools offer memory features, but says those can be difficult to read or edit and may be separate across tools. Those observations describe his motivation; the article does not compare specific products or independently assess their memory features.

Meyer’s alternative is a folder of text files with an index and scripts for retrieval and updates. Git records changes to the files over time, according to his description. The accompanying explainer presents nine views of the system, though the available source text only begins describing them and does not provide enough detail to characterize all nine. The diagrams are described as generic illustrations rather than screenshots.

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

— Thorsten Meyer, describing his system

How the Recall Score Was Measured

The article does not explain how the 97.4% score was calculated, how many questions were tested, what counted as the right page or whether the test included questions from everyday use. It also gives no independent evaluation. The score should therefore be read as a figure reported by the system’s creator, not as a verified benchmark or a prediction of performance in other settings.

Other operational details remain unspecified, including how the scripts handle questions that match several pages, fail to find a useful page or encounter incorrect or outdated notes. The source presents the system as personal and single-owner; it does not describe access controls, team workflows, backup arrangements or the safeguards used when agents write new information.

Testing the Personal Memory Approach

Meyer’s published account and explainer provide a description of the system, but the source gives no announced release, public implementation or independent evaluation schedule. A next step for readers assessing the approach would be to examine the scripts and files, if available, and test recall against a defined set of questions with known answers. Until those details are published, the reported page counts and recall score remain the author’s snapshot of his own setup as of Nov. 21, 2026.

Key Questions

What is Thorsten Meyer’s Second Brain?

It is a personal knowledge base that Meyer says he and his AI agents share. It consists of markdown pages, an index file and scripts for finding and adding information.

How do the agents find a memory?

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

How large was the system on Nov. 21, 2026?

Meyer reported 1,278 pages, 1,761 links and 1,363 recalls as of that date. The figures come from his own account.

Is the 97.4% recall score independently verified?

The source provides no independent evaluation or testing method. Meyer describes it as the rate at which the first page selected is correct, but readers cannot assess the score’s scope from the published details.

Is the system built for teams?

Meyer describes it as a single-owner setup. The source does not explain team permissions, collaboration or how competing edits would be managed.

Source: Thorsten Meyer AI

You May Also Like

The Impact Of Anthropic’s Claude AI Outage On Multiple Platforms

Anthropic experienced a major outage on August 16, affecting Claude.ai, Claude Code, and Claude Cowork, with services restored within 42 minutes.

How Meta Is Changing AI Development With The Muse Spark 1.2 Release

Meta launches Muse Spark 1.2 and Muse Code, co-trained for improved coding and tool use, intensifying competition in AI coding tools and agent-based models.

Understanding Anthropic’s $965B Series H: The Compute Revolution

Anthropic’s Series H puts compute capacity, chip supply and power at the center of the Claude growth story.

Discover The Best AI-Integrated Mobile Workstations Of 2026

Discover the top AI-enabled mobile workstations of 2026, featuring powerful GPUs, high RAM, and advanced displays for demanding professional tasks.