🔍 Read the full analysis: The Shared Memory That Brings My AI Agents Together on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Thorsten Meyer says he built a shared “Second Brain” so his AI agents can use the same long-term memory across sessions. The system stores information in Markdown files tracked with Git and reports a 97.4% top-one recall score as of November 21, 2026; the figures and performance measure come from Meyer, not an independent evaluation.
Thorsten Meyer has described a shared memory system that lets the AI agents he uses retrieve and add information across sessions, addressing a recurring limitation: an agent may lose access to the context of a past session when that session ends. Meyer’s “Second Brain” uses Markdown files tracked in Git rather than a cloud service or vector database, and he says it held 1,278 pages as of November 21, 2026.
Meyer’s account, published on ThorstenMeyerAI.com on November 21, 2026, presents the system as a personal knowledge base shared by his agents. It is intended to hold details about his work, decisions, projects and prior approaches so that an agent can consult the same information instead of asking him to repeat it in each new session. Meyer says any agent can look up information in the repository and add new facts to it.
The memory is organized as plain-text Markdown pages, with each page covering a cluster of related facts. A single file, INDEX.md, lists the pages, one line per memory. When an agent asks a question, a script called recall.py scores those index entries, opens the page it ranks highest and returns the section it considers relevant, along with the source page. Meyer says the script searches the index before opening a page rather than scanning the full collection.
A second script, remember.py, adds a page, its index entry and a log entry in one operation. Meyer says this approach is meant to prevent the index from falling out of sync with the stored notes. Git records changes over time. The system also has nine views, including tools for reviewing new session captures, tracing edits, exploring page links and monitoring the agents and services that use the memory.
The setup addresses a practical problem for people who rely on AI agents over time: separate sessions may not share the same record of a person’s preferences, prior decisions or project history. A common store can give multiple agents access to the same information, reducing the need to supply that context repeatedly. It also gives the user a place to inspect and correct what the agents retrieve.
Meyer’s choice of readable files makes the approach different from memory features that are difficult for users to inspect directly. Source-linked answers can help a user check which page informed a response, while Git offers a history of edits. Those features may make errors easier to trace, although they do not guarantee that the system selects the right page or that stored information is accurate.
The model has limits. Meyer describes a system built for one owner, not a shared team workflow, and maintaining useful pages and an accurate index requires ongoing organization. His report offers an example of one design, not evidence that the method will work equally well for other users or agent platforms.
Products Worth Considering

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

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

LG UltraGear 34G600A-B 34-Inch Curved WQHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How the Memory Is Organized
Meyer frames the project as an alternative to starting each agent session from scratch. An agent’s immediate context is what it can access during a session; when that session ends, the same information may not carry over. Some tools offer their own memory features, but Meyer says he wanted one store he could open, edit and use across the agents he runs.
In addition to the index, recall and write scripts, the system includes views for day-to-day oversight. Meyer describes a morning brief, a triage process for deciding which session captures become lasting memories, a chronicle of changes and a map of links between pages. The “Live” and “Desk” views show activity and workflows, while “Fleet” monitors the services and machines involved. These are features Meyer reports; the source does not provide an independent technical review.
As of November 21, 2026, Meyer reports 1,278 pages and 1,761 links. He also reports 1,363 recalls and a 97.4% top-one score, which he defines as how often the first page selected is the right one. The article does not specify the evaluation method, the number or type of questions tested, or an external comparison, so the score should be treated as the system owner’s reported measure.
““One memory, shared by every agent I run, that I can open and read myself.””
— Thorsten Meyer, describing the system on ThorstenMeyerAI.com
What the Reported Score Shows
The reported 97.4% top-one score is not independently verified in the source material. Meyer does not state how many test queries were used, how “the right” page was judged, or whether the measure was evaluated on a held-out set. It is also unclear how performance changes as the repository grows or as agents ask questions that span multiple pages.
The article does not give a detailed account of access controls, backups, privacy safeguards or how conflicting and outdated facts are handled. Meyer says personal information shown in the video and frames is protected, but that statement does not establish the system’s broader security properties. The source also does not report a team deployment or comparisons against other memory methods.
How Meyer Plans to Use It
Meyer’s article describes an operating system rather than announcing a product launch or a third-party release. He outlines daily and ongoing processes: reviewing session captures, deciding which details deserve permanent storage, checking changes and using shared memory during agent work. The source does not give a timetable for a public release, a team version or an independent performance test.
The next useful evidence would be a fuller account of how recall is measured, including the test set and the consequences of selecting the wrong page. For readers considering a similar setup, Meyer’s description provides a working example, while leaving the practical questions of maintenance, privacy and performance for different users unresolved.
Key Questions
What is Meyer’s Second Brain?
It is a personal knowledge base that Meyer says he shares with the AI agents he uses. Agents can retrieve information from it and add new facts.
How is the memory stored?
Meyer says it consists of Markdown files tracked in Git, with an INDEX.md file listing the memories. He says the system does not use a cloud service or vector database.
What does the 97.4% score mean?
Meyer defines it as the rate at which the first page selected by the system is the right one, calling it a top-one recall score. The source does not describe the test method or provide independent verification.
Can a team use the system?
Meyer characterizes the current setup as built for one owner, not a team. The source does not describe a team version or shared access controls.
What remains unknown about the system?
The source does not explain how the reported score was tested, how the system handles conflicting or outdated facts, or what security and access safeguards it uses. It also gives no public release plan.
Source: ThorstenMeyerAI.com
