🔍 Read the full analysis: How I Set Up Shared Memory For My AI Agents on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Thorsten Meyer says he built a shared knowledge base so his AI agents can retrieve and add information across sessions. The system uses Markdown files, a single index and scripts for recall and storage; its reported November 21, 2026 figures include 1,278 pages and a 97.4% top-one recall score.
Thorsten Meyer has described a shared long-term memory system for the AI agents he runs, built from Markdown files, a Git repository and a single index. In an article published November 21, 2026, Meyer said the system held 1,278 pages and recorded a 97.4% top-one recall score, figures he presents as measures of its scale and retrieval performance.
Meyer calls the knowledge base his Second Brain. His stated goal is to let multiple agents access the same information about his work, including past decisions, project details and approaches that did not work. Agents can retrieve existing information and add new material, rather than relying on separate memories tied to individual tools or sessions.
The underlying records are plain Markdown files stored in Git, with each page covering a cluster of related facts. A file named INDEX.md lists each memory in a single line. Meyer says a script called recall.py scores those index entries first, then opens the best-matching page and returns the relevant section with its source. The system does not use a vector database or cloud service, according to his account.
For adding information, Meyer says remember.py writes a memory page, its index entry and a log entry together. He describes that as a way to keep the catalogue aligned with the pages. His reported system totals also include 1,761 links and 1,363 recalls as of November 21. He defines top-one recall as how often the first page selected is the right one, but the article does not provide an independent evaluation or further detail on the test method.
The system addresses a practical limitation of session-based AI work: information available to an agent can be lost when a session ends, leaving users to repeat context. A shared store can give different agents access to the same project notes and decisions, while keeping the records in a format a person can inspect and edit.
That transparency is central to Meyer’s design. He says recalled answers identify the page they came from, allowing him to check the supporting record and correct errors at the source. For readers considering a similar setup, the trade-off is clear: text files and Git offer portability and visibility, but the system depends on careful indexing and maintenance. Meyer also says his setup is designed for one owner, not a team, so it should not be treated as a ready-made collaborative product.
Popular right now

Dyson V9 Motorbar Cordless Vacuum

KitchenAid 5.5-Quart Bowl-Lift Stand Mixer, Porcelain White

LG OLED evo C4 55-Inch 4K Smart OLED TV (OLED55C4PUA) with 144Hz Refresh Rate, HDMI Cable, 24-Month Protection Plan, and Streaming Bundle
As an affiliate, we earn on qualifying purchases.
From Session Context to Persistent Files
Meyer frames the project around the difference between an agent’s active context and durable memory. An agent can use information available during a session, he writes, but that context does not automatically carry forward after the session ends. He says some tools offer their own memory features, but he wanted one record he could access across the agents he uses and audit himself.
The published account describes a set of operational views around the files, including a map of links, a daily brief, a triage process for deciding what to retain, a change log and a console for monitoring services. These views organize and display the system; Meyer says the core memory itself remains the pages and index. He also says a video accompanying the article masks personal information shown in its frames.
“One memory, shared by every agent I run, that I can open and read myself.”
— Thorsten Meyer, describing the system in his November 21 article
How the Recall Score Was Measured
The figures are Meyer’s own reported measurements. His article does not explain the evaluation set, the number of test questions, how a correct page was determined or whether the score was measured over a particular period. It also gives no comparison baseline, so the 97.4% figure cannot be used to establish how the system performs relative to other memory methods.
The source account describes a personal setup, not a product assessment or a team deployment. It does not specify security controls, backup arrangements, access permissions for multiple users or the amount of ongoing work needed to maintain the index. Those details matter for anyone considering storing sensitive or shared organizational information in a similar system.
What Meyer Says the System Tracks
Meyer’s article presents the system as an ongoing personal tool, with views for reviewing new captures, tracking changes and monitoring the services and machines it uses. Its described workflow moves information from agent sessions through triage and into indexed pages, while a daily brief summarizes recent activity.
The published source does not announce a product release, outside evaluation or a planned team version. The next useful evidence would be additional detail on the recall test, maintenance demands and how the design handles access and security beyond a single owner. Until then, the reported figures describe Meyer’s system as of November 21, 2026, rather than independently verified performance.
Key Questions
What is Meyer’s Second Brain?
It is a shared knowledge base that Meyer says his AI agents can search and update. Its records are Markdown pages stored in Git, alongside a single index file.
How does an agent find a memory?
According to Meyer, recall.py scores entries in the index first, opens the best-matching page and returns the relevant section with its source.
What do the reported figures mean?
Meyer reported 1,278 pages, 1,761 links and 1,363 recalls as of November 21, 2026. He says the 97.4% top-one score measures how often the first page selected is the right one; the article does not give the test method or a comparison baseline.
Is the system built for teams?
No team deployment is described. Meyer says the setup is built for one owner, and the article does not detail shared-user permissions or collaborative controls.
Does the system use a cloud service or vector database?
Meyer says it uses neither. He describes the core as plain text files, Git and an index, with small scripts for recalling and adding information.
Source: ThorstenMeyerAI.com
