🔍 Read the full analysis: Shared Memory For AI Agents: The Second Brain Concept on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Thorsten Meyer describes a personal “Second Brain” that lets his AI agents share a long-term memory built from Markdown files and a searchable index. He reports 1,278 pages and a 97.4% top-1 recall score as of Nov. 21, 2026; the figures and performance measure come from his own account, not an independent evaluation.
Thorsten Meyer has described a shared long-term memory system for the AI agents he runs, built from Markdown files, a single index and Git rather than a vector database or cloud service. In an article published Nov. 21, 2026, he said the system held 1,278 pages and had a reported 97.4% top-1 recall score, offering a personal example of how agents can reuse information across sessions instead of starting from scratch.
Meyer calls the system his Second Brain. He says it gives his agents and himself access to the same knowledge base: an agent can look up information about his work, while he can inspect or correct the underlying files. Each Markdown page stores a cluster of related facts, such as material about a project or a decision. Git records changes to those files over time.
A file called INDEX.md lists each memory in a single line. When an agent needs information, a script named recall.py scores those index lines, opens the page it ranks highest and returns the relevant section with its source. Meyer says the system does not search every page during that step. Its reported 97.4% top-1 score measures how often the first page selected is the right one, according to his account.
New information is added through remember.py, which writes the memory page, its index entry and a log entry together. Meyer says this is meant to keep the index aligned with the files. He also describes nine monitoring and management views, including tools for reviewing new captures, tracking changes, exploring links between pages and watching agents recall or add information. The article lists 1,761 links and 1,363 recalls as of Nov. 21, 2026.
The system addresses a practical limitation of many AI-agent workflows: information available in one session may not be available in the next. A shared store can let different agents reuse decisions, project details and prior lessons without requiring the user to repeat them each time. Meyer’s design also makes the stored information readable and editable outside an agent interface, and its source-linked responses give the owner a way to check what an answer relies on.
That approach trades some convenience for visibility and control. Plain files can be inspected with ordinary tools, and Git provides a history of changes. But Meyer describes a system designed for one owner, not a team. Keeping records current, deciding what should become a lasting memory and managing access remain human and operational responsibilities. His account is a case study, not evidence that the same setup will perform equally well for other users or workloads.
Popular right now

De'Longhi Magnifica Evo Automatic Espresso Machine with Milk Frother, Silver

Bose QuietComfort Wireless Bluetooth Noise-Cancelling Headphones, Ice Blue

Samsung 65-Inch Q8F QLED 4K UHD Smart TV (2025)
As an affiliate, we earn on qualifying purchases.
From Session Context to Persistent Files
AI agents generally use information available to them during a session, including what fits in their current context window. Meyer says that when a session ends, that immediate context is gone, and he found himself repeatedly explaining the same work and decisions to agents. He says some tools provide memory features, but he wanted a shared record he could read, audit and correct across the agents he uses.
His response is a deliberately simple storage structure: Markdown pages, one index and Git. The index is central because the recall script relies on its short descriptions to select a page. The write script is designed to update the page, catalogue and log in one operation. Meyer’s article presents this as a working personal system and provides figures dated Nov. 21, 2026; it does not describe a controlled comparison with other memory approaches.
“One memory, shared by every agent I run, that I can open and read myself.”
— Thorsten Meyer, describing the system in his article
How the Recall Score Was Measured
The reported 97.4% top-1 recall score is Meyer’s own measure. The article does not specify the test set, number of test questions, evaluation method or whether an outside party checked the result. It also does not compare the system’s accuracy or speed with other memory tools. The figure should be read as a result reported by the system’s creator, not as an independently verified benchmark.
The article describes the setup as personal and says it uses no cloud service, but does not provide a detailed security assessment or explain how the scripts handle conflicting, outdated or sensitive memories. It is also unclear how the approach would work with multiple users, larger collections or agents that need to search across several relevant pages rather than the single top-ranked result.
Adoption Depends on Testing and Upkeep
Meyer’s article does not announce a product launch or give a public release schedule. It presents the Second Brain as a system he built for his own work and explains its structure and reported performance. Readers considering a similar setup would need to assess whether file-based storage, Git history and a single index fit their agents and information practices.
The next useful evidence would be fuller documentation of the recall evaluation, including its test questions and failure cases, as well as practical results from other users or teams. Until then, the reported figures describe the system as Meyer says it stood on Nov. 21, 2026; its reliability in different settings remains unestablished.
Key Questions
What is the Second Brain?
It is Meyer’s personal knowledge base shared with the AI agents he runs. It stores information in Markdown files and uses an index to help agents find relevant pages.
How does an agent retrieve information?
According to Meyer, the recall.py script scores the lines in the index, opens the highest-ranked page and returns the section it considers relevant, with a source reference.
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. These are figures from his article.
Is the 97.4% recall score independently verified?
The source article does not provide an independent evaluation or enough information about the test method to verify the result. The 97.4% figure is reported by Meyer.
Is the system designed for teams?
Meyer describes it as a setup built for one owner, not a team. The article does not explain how shared permissions or team workflows would be handled.
Source: ThorstenMeyerAI.com
