Can One Memory Help AI Agents Collaborate?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Can One Memory Help AI Agents Collaborate? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Thorsten Meyer says he built a shared memory system for his AI agents using Markdown files, a searchable index and scripts that retrieve or save information. As of Nov. 21, 2026, he reported 1,278 pages and a 97.4% top-one recall score; those figures are his system’s own measurements, and the setup is designed for one owner rather than a team.

Thorsten Meyer has described a shared memory system that lets the AI agents he runs consult and update the same collection of information, rather than starting each session without knowledge of past work. In an article published Nov. 21, 2026, Meyer said the system contains 1,278 pages and has a 97.4% top-one recall score—a measure he defines as how often its first selected page is the right one.

Meyer calls the collection his Second Brain. Its contents are plain Markdown files stored in Git, with each page covering a cluster of related facts, such as a project or a decision. A single file, INDEX.md, lists the pages. Meyer says the system uses neither a vector database nor a cloud service, and that the files can be opened and edited with ordinary text editors.

When an agent needs information, a script called recall.py scores the index entries first, then opens the best-matching page and returns the relevant section with its source. Meyer says this lets him check which page informed an answer. To add information, remember.py writes the page, an index entry and a log entry together. He says this process is intended to keep the index and stored notes aligned.

As of Nov. 21, Meyer reported 1,278 pages, 1,761 links and 1,363 recalls. The article also describes nine views for monitoring and using the system, including a map of connections, a daily brief, a triage process for session notes and a chronicle of changes. Meyer presents the project as a working setup for his own agents, not as a standardized product or a controlled comparison with other memory systems.

At a glance
reportWhen: Published Nov. 21, 2026; system figures…
The developmentThorsten Meyer published a description of a personal, shared memory system for his AI agents, including its file-based design and reported usage metrics.

Shared Memory Across Agent Sessions

A shared store can address a practical limitation of session-based agents: information from one interaction may not be available in the next. With a common memory, Meyer’s agents can refer to recorded project details and decisions without asking him to repeat them each time. The arrangement also gives him a single place to review and correct stored information, rather than relying on separate memory features tied to individual tools.

The design makes a trade-off between simplicity and scale. Readable files and Git history make the stored material inspectable and trackable, while index quality and careful organization become important to retrieval. Meyer says the setup is built for one owner, not a team; the article does not establish how well it would work with multiple people editing, sharing permissions or conflicting versions of information.

The reported recall score offers a snapshot of this particular system, but it is not an independent benchmark. Meyer does not provide a testing protocol or comparison baseline in the material supplied. Readers should treat the 97.4% figure as the author’s own metric, not evidence that the approach will perform similarly across agents, tasks or organizations.

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

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

Vitamix 5-Speed Immersion Blender, 625-Watt, Stainless Steel/Black

Vitamix 5-Speed Immersion Blender, 625-Watt, Stainless Steel/Black

As an affiliate, we earn on qualifying purchases.

From Session Notes to Indexed Files

Meyer’s stated motivation is that an agent’s working context is limited to what it can access during a session. When a session ends, he says, that context is gone. He also argues that memory features offered by separate tools can be difficult to inspect or audit and may not be shared across his agents. His system is intended to keep one accessible record that the agents can consult and update.

The workflow has two central operations: recall, which searches index lines before opening a page, and remember, which stores a new fact along with its index and log entries. Meyer says a triage process reviews information captured from sessions and decides what should become a lasting memory, merge with an existing page or be dropped. Git records changes over time, while the system’s views provide ways to inspect links, activity and service status.

The article documents Meyer’s own implementation and includes descriptions of an accompanying video and interface views. It does not identify an external study, independent audit or release of a general-purpose product. The system’s reported scale and accuracy should be understood within that personal-use account.

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

— Thorsten Meyer, describing the system in his article

Limits of the Reported Accuracy

The article gives no details about how the 97.4% top-one score was calculated, including the number and type of test questions, who judged whether the selected page was correct, or how the result changes across different tasks. It also does not state a comparison baseline, so the figure cannot establish whether the system performs better than other retrieval methods.

It remains unclear how the design handles access controls, sensitive information, conflicting updates or use by multiple people. Meyer describes the system as a personal setup and says personal data in the video and frames is protected, but the supplied material does not detail a broader security review or independent verification of the reported counts. The metrics are therefore author-reported figures, not independently confirmed results.

How Meyer Plans to Use the System

Meyer’s article describes the system’s current workflow and monitoring views but does not announce a product launch, public release date or external evaluation. The next steps described within the setup include reviewing session captures through triage, maintaining the index as new memories are added and using its daily and health views to track changes.

For readers considering a similar approach, the account leaves practical questions open: how much manual review the system requires over time, whether retrieval quality holds as the collection grows, and what changes would be needed to support a team. No independent test or team deployment is reported in the source material, so those questions remain unresolved.

Key Questions

What is Meyer’s Second Brain?

It is a shared knowledge base for the AI agents Meyer runs, built from Markdown files and an index. Agents can retrieve information from it and add new memories.

How does an agent find information?

Meyer says the recall.py script scores entries in the index, opens the highest-scoring page and returns the relevant section with its source.

What does the 97.4% figure measure?

Meyer defines top-one recall as how often the first page selected by the system is the right one. The article does not give a test protocol or an independent verification, so the figure is an author-reported measure.

Can a team use the same setup?

The system is described as a single-owner setup. Meyer does not report a team deployment or explain how permissions and simultaneous edits would be managed.

Does the system use a vector database or cloud service?

Meyer says it uses plain Markdown files stored in Git, with an index and scripts, and that it does not use a vector database or cloud service.

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 Luddites Weren’t Simply Afraid Of Machines

A historical account argues that Nottingham’s frame breakers targeted low pay and poor-quality production, not machinery itself.

Technology Operations Signal Monitor: How Google Helped Destroy Adoption Of RSS Feeds (2023)

New analysis shows Google’s platform and tooling changes contributed to the decline of RSS feed usage, impacting small software companies’ product decisions.

The UGREEN NASync DXP6800 Pro: Our 6-Bay Choice For AI

A small AI publisher bought UGREEN’s six-bay NAS after backup-sync problems. The device is delivered but not yet configured or tested.

Find Your Fit: 10 Best Ultrawide Monitors For Work And Gaming In 2027

Compare 10 ultrawide monitors by resolution, refresh rate, connectivity and size, with picks for work, gaming and large desktop setups.