How My Agents Remember More With One Shared Brain
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: How My Agents Remember More With One Shared Brain on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

Thorsten Meyer says he built a shared, Git-backed memory system so his AI agents can use the same accumulated knowledge across sessions. As of November 21, 2026, it contained 1,278 pages; Meyer reports a 97.4% top-1 recall score, though the article does not provide an independent evaluation or testing details.

Thorsten Meyer has described a shared long-term memory system for the AI agents he runs, using Markdown files and a single index so different agents can draw on the same information between sessions. Meyer says the system held 1,278 pages as of November 21, 2026, and reports a 97.4% score on its top-1 recall measure; the figures and performance assessment come from Meyer, and the article provides no independent verification.

Meyer’s system, which he calls his Second Brain, is intended to address a recurring limitation of agent sessions: information available in an agent’s context window is not necessarily available after that session ends. Rather than re-explain project details and previous decisions to each agent, Meyer says he lets them consult a common store and add information to it.

The memory is made up of plain Markdown files kept in Git. Each page contains a cluster of related facts, while an index file called INDEX.md lists the pages, one line at a time. Meyer says a recall script, recall.py, scores those index entries before opening the best-matching page and returning the relevant section with its source. He says the design uses no vector database or cloud service.

Adding information is handled by a second script, remember.py. According to Meyer, it writes the page, an index entry and a log line together. He says this is intended to prevent the index from becoming out of step with the underlying notes and to preserve a chronological record of changes. The published account reports 1,761 links and 1,363 recalls alongside the 1,278 pages, all counted as of November 21, 2026.

At a glance
reportWhen: Published November 21, 2026; system fig…
The developmentThorsten Meyer published a description of a plain-text knowledge base that lets his AI agents share and retrieve long-term memories.

A Shared Record Across Agent Sessions

The system offers one approach to a practical issue for people using several AI agents: knowledge can be scattered across tools and sessions, making it harder to reuse earlier decisions or check where an answer came from. A common, readable store could make that information available to more than one agent while giving its owner a direct way to inspect and correct it.

Meyer’s emphasis on source-linked recall is relevant to oversight. If an agent returns a fact alongside the page it came from, a user can check the underlying note rather than treating the answer as self-explanatory. Keeping files in Git also provides a history of edits, according to Meyer. These are design features described by the author, not evidence that the system prevents errors or guarantees accurate answers.

The trade-off is that the method depends on an owner maintaining useful pages and index entries. Meyer says the setup is built for one owner rather than a team. Its fit for collaborative use, and how it compares with built-in memory tools, are not established in the account.

Products Worth Considering

Apple iMac MK472LL/A 27‑Inch Retina 5K Desktop (3.2 GHz Intel Core i5, 8 GB DDR3, 1 TB, macOS)

Apple iMac MK472LL/A 27‑Inch Retina 5K Desktop (3.2 GHz Intel Core i5, 8 GB DDR3, 1 TB, macOS)

As an affiliate, we earn on qualifying purchases.

From Session Context to Stored Notes

An AI agent’s context window contains the information it can use during a given session. Meyer’s account starts from the problem that details from one session may not carry into the next, so users can find themselves repeating project background or explaining which approaches have already been tried. He says existing memory features can be difficult to read, edit or audit, and may be tied to individual tools.

His alternative is a local collection of text files, an index and scripts for retrieval and writing. The article also describes nine views for monitoring the system, including a map of connections, a daily brief, a triage view for deciding which session captures to keep, and logs of changes. These views are interfaces around the file-based store; Meyer says the files and index are sufficient for the underlying memory to operate.

““One memory, shared by every agent I run, that I can open and read myself.””

— Thorsten Meyer, describing the system

How Recall Performance Was Measured

Meyer reports a 97.4% top-1 recall score, defining the measure as how often the first page selected by the system is the right one. The article does not specify how many queries were tested, how “the right one” was judged, whether the test set was independent, or how the result was recorded. The score should therefore be understood as a figure reported by the system’s author, not an independently validated benchmark.

The account also does not detail safeguards for inaccurate, outdated or sensitive information, or describe how multiple users would manage permissions and disagreements. Meyer says personal data in the accompanying video and frames is protected, but provides no further technical detail about those protections. The reported counts describe the system on November 21, 2026, rather than a continuing measure of its size or performance.

Further Testing and Team Use

Meyer’s article points readers to an explainer video showing the system’s views and operations. It does not announce a product release, public repository, roadmap or formal evaluation. The next useful evidence would include a clearer account of the recall test, examples of errors and corrections, and details about how the system handles sensitive data.

For organizations considering a similar approach, questions remain about shared editing, access controls and maintenance at team scale. Meyer’s description presents a working personal setup; whether its file-and-index method can support those needs is not established. No further milestone or release date is specified.

Key Questions

What is Meyer’s Second Brain?

It is a shared knowledge base that Meyer says his AI agents can consult and update. It consists of Markdown pages, an index and scripts for retrieving and recording information.

How does an agent find a memory?

According to Meyer, recall.py scores lines in the index, opens the best-matching page and returns the relevant section with its source. The described process checks the index before opening a page.

What does the 97.4% figure mean?

Meyer calls it the system’s top-1 recall score: how often the first page selected is the right one. The article does not give test size, methodology or independent validation, so the figure cannot be compared confidently with other systems.

Does the system require a vector database or cloud service?

Meyer says it uses plain Markdown files in Git and does not use a vector database or cloud service. The account does not give a full deployment or security specification.

Can a team use the same setup?

Meyer describes the current arrangement as designed for one owner, not a team. The article does not explain how collaborative access, permissions or shared editing would work.

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

The Quiet Audit: 55–75% of Your Week Is on Thin Ice. Here’s Which Part.

A detailed analysis shows 55-75% of knowledge workers’ time is on thin ice, with most of it being routine, performative, or automatable tasks.

The Future Of Artificial Intelligence Hinges On Pre-Designed Hardware

AI hardware is shifting from general-purpose GPUs to purpose-built chips optimized for inference workloads, driven by thermal, memory, and specialization advances.

US Data Centers Face A Four-Part Test On Power Capacity

Rymvard published illustrative scenarios showing how grid delays, curtailment, cooling and tariffs can limit usable data center capacity.

Fable and Mythos: How Anthropic Shipped Its Most Powerful Model to Everyone

Anthropic launches Fable 5, a highly capable AI model with safety features allowing broad access. Mythos 5 remains restricted for safety reasons.