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

🔍 Read the full analysis: Why My AI Agents Need A Shared Second Brain on ThorstenMeyerAI.com

TL;DR

Thorsten Meyer says he built a shared Second Brain so his AI agents can use the same long-term work memory across sessions. The system uses Markdown files, a Git history and an index-based retrieval script; Meyer reports 1,278 pages and a 97.4% top-1 recall score as of Nov. 21, 2026. The account describes one owner’s setup, not an independently evaluated or team-tested product.

Thorsten Meyer says he built a shared long-term memory for the AI agents he uses, allowing them to consult and add to the same collection of work notes across sessions. In an account published Nov. 21, 2026, he described a system based on Markdown files, Git and a searchable index, and reported that it contained 1,278 pages with a 97.4% top-1 recall score.

Meyer’s central problem is that an agent’s working context does not automatically carry over when a session ends. He says he was repeatedly explaining projects, previous decisions and failed approaches to agents, and chose to keep that information in a separate knowledge base he could inspect himself. The agents can look up information in the collection and write new material to it.

Each memory is stored in a plain-text Markdown page, with pages organized around clusters of related facts, such as a project or decision. A single file, INDEX.md, lists the memories. Meyer says a script called recall.py scores index entries, opens the best-matching page and returns the relevant section with its source. The approach does not use a vector database or cloud service, according to his account.

For new information, Meyer describes a second script, remember.py, that writes a page, its index entry and a log entry in one operation. He says this is intended to keep the catalogue synchronized with the files and provide a record of changes. As of Nov. 21, the system had 1,278 pages, 1,761 links and 1,363 recalls. Meyer’s reported top-1 score measures how often the first page selected by the system is the right one; he gives the result as 97.4%.

At a glance
reportWhen: Published November 21, 2026; reported s…
The developmentThorsten Meyer published an account of a shared, file-based memory system he built for the AI agents he uses.

Shared Memory Across Agent Sessions

The system addresses a practical limitation of using several AI agents: information entered into one session may not be available to another. A shared, owner-controlled repository gives agents a common reference point, while letting the person using them inspect or correct the underlying material. Meyer argues that showing the source page alongside a retrieved answer makes it easier to check how the agent reached that information.

The design also makes a trade-off visible. Plain files and Git history can be read and moved with familiar tools, but they require consistent organization and maintenance. Meyer says the setup is built for one owner rather than a team. His account therefore illustrates one approach to agent memory, not evidence that it will work equally well for collaborative use or other users’ workloads.

Ninja Foodi 10-Quart DualZone Smart XL Air Fryer, 6-in-1, Black, DZ550

Ninja Foodi 10-Quart DualZone Smart XL Air Fryer, 6-in-1, Black, DZ550

Sony WH-1000XM6 Wireless Noise-Cancelling Headphones – Midnight Blue

Sony WH-1000XM6 Wireless Noise-Cancelling Headphones – Midnight Blue

Philips 2300 Series Fully Automatic Espresso Machine with LatteGo

Philips 2300 Series Fully Automatic Espresso Machine with LatteGo

As an affiliate, we earn on qualifying purchases.

From Session Context to File-Based Recall

An AI agent responds using information available in its current context, alongside any external tools or memory features it can access. Meyer’s account focuses on what happens between sessions: he wanted information about his work to remain available without relying on separate, opaque memory stores for each tool. His proposed answer is a repository shared by every agent he runs.

The system includes nine named views, according to Meyer. They cover connections between projects and tools, daily summaries, review of captured session notes, change history, page links and live agent activity. He also describes a control-desk view for daily operations and a fleet console for monitoring services and machines. These are ways to inspect and manage the underlying collection; the core storage remains files and an index.

Meyer says Git records the repository’s change history, while the system’s log records additions and changes in sequence. He also says personal information shown in the accompanying video and images is protected or masked. The published description is a first-person account by the system’s creator, rather than an outside audit.

„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-1 score is Meyer’s own measure. His account does not specify the size or composition of the test set, how a correct first result was judged, or whether an independent party evaluated the result. It also does not provide a comparison with other retrieval systems, so the figure should not be read as a benchmark against alternatives.

The article describes the architecture and the number of pages, links and recalls, but does not establish how the setup performs for other people, larger teams or different types of work. It is also not clear how the system handles conflicting or outdated memories beyond the described review and logging tools. Meyer notes that the arrangement is designed for one owner, leaving its suitability for shared company use unresolved.

Testing the Setup Beyond One Owner

Meyer’s account does not announce a commercial release, public product launch or independent evaluation. The immediate next step for readers interested in the approach is to review the described workflow and decide whether maintaining plain-text files, a central index and a change log fits their own agent use. Any broader claims about reliability or team adoption would require evidence beyond the figures Meyer reports for his personal system.

Further information that would clarify its performance includes a published testing method for the top-1 score, examples of retrieval errors and details of how memories are reviewed as projects change. Until then, the reported results describe the system as it stood on Nov. 21, 2026, rather than a guarantee of future accuracy.

Key Questions

What is Meyer’s Second Brain?

It is a shared knowledge base that Meyer says he uses with his AI agents. It stores work-related information in Markdown files and lets agents retrieve or add to those records.

How does an agent find a memory?

According to Meyer, recall.py scores entries in an index, opens the best-matching page and returns the relevant section with its source.

What figures did Meyer report?

As of Nov. 21, 2026, Meyer reported 1,278 pages, 1,761 links and 1,363 recalls, along with a 97.4% top-1 score. The score is his own measure, and the account does not give an independent test method.

Is the system designed for teams?

Meyer says it is built for one owner, not a team. The account does not show how it would handle collaboration across multiple people.

Does the system use a cloud service or vector database?

Meyer says the core system uses plain Markdown files in Git and an index, with no vector database or cloud service.

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

Did The Tampered Claude Code Update Reach Your Cursor Setup?

Cursor says a modified Claude Code 2.1.313 was available for about 11 hours, potentially affecting up to 5,000 users. Here’s what to check.

Agentic Loop Failure Modes: A Production Taxonomy at the End of Year One

A new taxonomy categorizes the failure modes of production agentic systems after their first year, aiding debugging and architectural decisions.

What’s Established In The Reports On Nvidia, Trump And The Spy Probe?

Nvidia confirmed a security inquiry involving Jensen Huang. Reports say directors discussed his future, but the company has announced no decision.

How Market Oversights Are Undermining AI Token Potential

Analysis of how market mispricing and overlooked demand in open-source AI models are undervaluing AI tokens, despite rising fundamental activity.