🔍 Read the full analysis: How One Memory Gives My AI Agents Shared Context on ThorstenMeyerAI.com
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
TL;DR
Thorsten Meyer says he built a shared “Second Brain” so his AI agents can consult and update the same long-term memory across sessions. The system uses Markdown files, Git and an index rather than a vector database or cloud service; Meyer reports 1,278 pages and a 97.4% top-one recall score as of Nov. 21, 2026. The figures and performance measure come from Meyer, and the account does not provide an independent evaluation.
Thorsten Meyer says he has built a shared long-term memory system that lets the AI agents he runs retrieve and add information across sessions. In an article published Nov. 21, he describes a setup of Markdown files kept in Git and a single index, reporting 1,278 memory pages and a 97.4% top-one recall score as of that date. The account is Meyer’s description of his own system; the performance figure is not independently evaluated in the material provided.
Meyer calls the system his Second Brain. He says it is meant to address a recurring limitation of agent sessions: an agent can use what is present in its active context window, but after a session ends that information may not carry over. Instead of having to explain work, decisions and failed approaches again, Meyer’s agents can consult a memory shared across the tools he runs.
The stored information is organized in plain Markdown files, with each page covering a cluster of related facts, such as a project or decision. Git records changes to the files. A separate file, INDEX.md, provides one line for each memory. Meyer says a script named recall.py scores those index lines, opens the highest-ranked page and returns only the section it judges relevant, along with the page source.
New information is added through remember.py, which writes the memory page, its index line and a log entry together. Meyer says the combined step is intended to stop the index and pages from falling out of sync. The system also has nine views, including a map of connections, a daily brief, a triage process for session captures and a live view of agent activity. Meyer reports 1,278 pages, 1,761 links and 1,363 recalls as of Nov. 21, 2026; the source does not define the time period covered by the recall count.
The design addresses a practical problem for people using multiple agents: information from one session or tool may not be available to another. A common, readable store gives Meyer’s agents a route to the same project details and decisions, while allowing him to inspect and correct the underlying files. That makes the system’s memory less dependent on a single agent interface, according to Meyer’s account.
Its source-linked recall is central to the approach. Meyer says an agent’s answer identifies the page it used, letting him check the record rather than treating the response as an unsupported recollection. The method also trades some automation for explicit organization: the index must describe each page well enough to help the recall script select it, and the system depends on a controlled process for adding and maintaining memories.
Meyer presents the setup as suitable for an individual owner, not as a ready-made team product. That distinction matters for organizations considering shared agent memory: access controls, collaboration, conflict handling and oversight would need answers beyond the personal workflow he describes. The reported recall score is a system-specific measure, not evidence that the same design will perform similarly on other collections or questions.
Popular right now

Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Black

Roborock Qrevo S Pro Robot Vacuum and Mop (2026)

KitchenAid Artisan Series 5-Quart Tilt-Head Stand Mixer, Juniper (KSM150PS)
As an affiliate, we earn on qualifying purchases.
Files, Git and an Index
Meyer’s design places the memory outside any one agent’s active conversation. He says there is no vector database and no cloud service in this setup: the memory is a folder of text files that can be opened in an editor. Git provides a history of changes, while the index gives the recall process a compact catalogue to search before opening a page.
The process has two stages: index first, files second. Rather than reading every memory page for each query, the script scores the index entries, selects a page and returns a relevant section. Meyer says the recall mechanism is deterministic, but the article does not provide its scoring formula or test procedure. The system’s nine views add ways to monitor memory health, connections, changes and agent activity, while the underlying store remains the files and index.
“I got tired of being their memory. So I built them one.”
— Thorsten Meyer, describing his motivation
How the Recall Score Was Measured
Meyer reports that the system’s “top-1” recall score is 97.4%, describing it as how often the first page selected is the right one. The article does not specify the number or type of queries in the evaluation, how “right” was judged, or whether the score was measured on a fixed test set. The figure should therefore be read as Meyer’s own system measure, not as an independently verified benchmark or a general measure of AI-agent accuracy.
The source also does not state how often the memory is updated, how errors are reviewed, or how the setup would handle multiple people editing it. It gives a count of 1,363 recalls but no time window for that count. Meyer says personal information in the accompanying video and frames is protected, but the material does not detail the protections used for the underlying files or the machines that run the system.
From Personal Setup to Wider Use
Meyer’s article presents the current system and points readers to an explainer video that walks through its views and workflow. It does not announce a launch date, public download, product release or planned expansion. The next visible development is therefore further documentation or a decision by Meyer to make the tools and methods available beyond his own setup; neither is confirmed in the source.
For readers assessing the approach, the practical next step is to examine the described workflow: how the index is maintained, what counts as a useful memory, how records are corrected and whether citations reliably lead to supporting information. Until the test method and team-use details are available, the page count and recall score describe Meyer’s reported experience, not a guarantee of results for other users.
Key Questions
What is Meyer’s Second Brain?
It is a shared long-term memory that Meyer says his AI agents can read and update across sessions. Its information is stored in Markdown files with an index.
How does an agent find a memory?
According to Meyer, recall.py first scores entries in the index, then opens the highest-ranked page and returns a relevant section with its source.
What does the reported 97.4% score mean?
Meyer describes it as the system’s top-one recall: how often the first page selected is the right one. The source does not explain the test set or evaluation method, and the result is not independently verified in the material provided.
Is the system designed for teams?
Meyer says the setup is built for one owner, not a team. The source does not describe team permissions, collaboration or conflict resolution.
Source: ThorstenMeyerAI.com
