🔍 Read the full analysis: The Role Of Shared Memory In My AI Agent Setup 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 the AI agents he runs can read and add to one long-term memory. The system uses Markdown files, Git and a searchable index rather than a vector database or cloud service; Meyer reports 1,278 pages and a 97.4% top-1 recall score as of Nov. 21, 2026.
Thorsten Meyer has described a shared long-term memory system for the AI agents he runs, built from Markdown files stored in Git and a single index that agents consult before retrieving details. Meyer says the system held 1,278 pages as of Nov. 21, 2026, and that its first-result recall measure was 97.4%; those figures are his reported measurements, not an independently verified evaluation.
Meyer calls the system his “Second Brain”. In his account, it addresses a recurring limitation of agent sessions: information available in one session may not carry over when that session ends. Rather than maintaining separate memories for each tool, he says he gives all his agents access to one knowledge base that he can also read and edit.
Each memory is kept in a plain Markdown file, with pages organized around a cluster of related facts, such as a project or decision. A file named INDEX.md lists the pages, one line per memory. Meyer says a deterministic script called recall.py scores those index lines, opens the best-matching page and returns the relevant section with its source, instead of searching every page.
New information is added using remember.py, which writes the page, an index entry and a log entry together. Meyer says this is intended to prevent the index from becoming inconsistent with the stored notes. Git records changes over time. He reports 1,761 links between pages and 1,363 recalls, alongside the page count and recall score, as of Nov. 21. The source does not provide a testing protocol or independent audit for those metrics.
The setup offers one answer to a practical problem for people who use multiple agents: repeatedly supplying the same background and trying to keep separate tool memories in sync. A common store can make information available across sessions and agents, while keeping the underlying material in a format its owner can inspect.
That inspectability is part of Meyer’s stated rationale. Because the system returns a source page with a recalled answer, he says he can check where the information came from and correct it at the source. The design also trades some convenience for manual organization and maintenance: useful retrieval depends on clear index descriptions and reliable handling of new facts.
Meyer frames the system as personal rather than team infrastructure. A folder of text files and Git may be accessible and portable, but the article does not establish how the approach would handle shared permissions, conflicting edits, sensitive information or a large organization’s governance needs. Its reported score should be read as a measure from one creator’s own setup, not a general benchmark for agent memory systems.
Products Worth Considering

Samsung Odyssey OLED G8 34-Inch Curved Ultra WQHD Gaming Monitor

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)

Jackery HomePower 3600 Plus Portable Power Station
As an affiliate, we earn on qualifying purchases.
From Session Context to Files
An AI agent can use information included in its active context, but the source article says that information does not automatically persist after a session ends. Meyer says he wanted a memory that was readable, editable and auditable, rather than one held in a service-specific store that he could not readily inspect. The article describes his own implementation; it does not compare it with particular commercial products or establish that those products all work the same way.
Beyond recall and storage, Meyer describes nine views for monitoring and organizing the system. These include a Brief summarizing recent sessions and new memories, a Triage view for deciding what to keep or merge, and a Chronicle that records changes. Other views show links between pages, live activity and the services or machines involved. Meyer says personal information in the accompanying video and still images is protected or masked.
““I got tired of being their memory. So I built them one.””
— Thorsten Meyer, describing his motivation
How the Recall Score Was Measured
The article reports a 97.4% top-1 recall score, defining top-1 as how often the first page selected is the right one. It does not describe the test set, how “right” was judged, how many queries were included, or whether the evaluation was independently repeated. The figure therefore cannot be directly compared with other memory systems from the information provided.
It is also unclear how the setup handles overlapping or outdated facts, access controls when multiple people use a memory, backup and recovery, or sensitive data beyond the described protections for video examples. Meyer says there is no cloud service in this design, but the article does not offer a security review or discuss how the system behaves under failures. The claimed ease of checking and moving files does not, by itself, establish reliability or privacy protections for every deployment.
Further Detail on Testing and Use
Meyer’s article points readers to an explainer video showing the system’s views and workflows. It does not announce a release schedule, a public software package or a plan to turn the personal setup into a team product. The next useful evidence would be more detail on the recall evaluation, including its query set and scoring method, along with documentation of how the scripts and memory files are maintained.
For now, the system should be understood as Meyer’s described implementation and a set of reported measurements at a specific date. Whether the approach works as well for other users will depend on their information, indexing practices and needs for collaboration and security.
Key Questions
What is Meyer’s Second Brain?
It is a personal knowledge base that Meyer says he and the AI agents he runs can both read and update. Its memories are stored as Markdown files with a shared index.
How does an agent find a memory?
According to Meyer, recall.py scores entries in the index, opens the best-matching page and returns the relevant section with its source.
What performance figures did Meyer report?
As of Nov. 21, 2026, Meyer reported 1,278 pages, 1,761 links and 1,363 recalls. He also reported a 97.4% top-1 recall score, but did not provide the testing details needed to compare it with other systems.
Is the system designed for teams?
Meyer describes it as a setup built for one owner, not a team. The article does not detail team permissions, collaborative editing or organizational controls.
Does the approach use a cloud service or vector database?
Meyer says this implementation uses local plain-text files and Git, with a single index, and no vector database or cloud service. The article does not provide an independent technical audit of the setup.
Source: ThorstenMeyerAI.com
