How My AI Agents Pass Context Through One Memory
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: How My AI Agents Pass Context Through One Memory on ThorstenMeyerAI.com

Before you orderOffer from Amazon

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
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 his AI agents use one shared memory made from Markdown files in Git and a single index. As of November 21, 2026, it held 1,278 pages and had a reported 97.4% top-1 recall score; the figures and performance measure come from Meyer, and the setup is designed for his personal use.

Thorsten Meyer has described a shared long-term memory that lets the AI agents he runs retrieve and add information across sessions, rather than starting each session without his work history. In an article published November 21, 2026, Meyer said the system uses Markdown files stored in Git, a single index and scripts for retrieval and updates; he reported 1,278 pages and a 97.4% top-1 recall score as of that date.

Meyer calls the system his Second Brain. Its pages store clusters of information, such as material about a project or a decision. A central file, INDEX.md, lists each page in a single line. According to Meyer, recall begins by scoring those index lines, then opens the best-matching page and returns the relevant section with its source. He says the setup does not use a vector database or cloud service.

New information is added through a script called remember.py, which writes the memory page, an index entry and a log entry together. Retrieval uses a separate script, recall.py. Meyer says this single-step process is intended to keep the index aligned with the pages and give him a record of changes. He reported 1,761 links and 1,363 recalls alongside the page count, though the article does not define the period covered by the recall total.

The system includes nine views for monitoring and managing the collection. Meyer describes tools for mapping links, reviewing new session captures, checking frequently used pages and watching services. He says the views expose how information moves from a session into a lasting page, and that personal information shown in the accompanying video and images is protected or masked.

At a glance
reportWhen: Published November 21, 2026; system fig…
The developmentThorsten Meyer published a description of a shared, file-based memory system that lets his AI agents retrieve and add information across sessions.

A Shared Record Across Agent Sessions

The approach addresses a practical limitation of session-based agents: when a session ends, its immediate context may not carry over to the next one. A common store can let separate agents refer to the same project decisions and prior work, reducing the need for a user to repeat information. That is the intended benefit Meyer describes, not an independently measured improvement in agent performance.

Keeping information in readable, version-controlled files also gives the owner a way to inspect sources, correct errors and review changes. This differs from memory features that may be tied to one product or stored in a format a user cannot readily edit. The trade-off is ongoing maintenance: the index must remain useful, pages need to be curated, and Git history alone does not resolve whether a stored fact is accurate or still current.

Meyer characterizes the system as built for one owner, not a team. That limits what can be concluded about its use in organizations, where permissions, simultaneous edits, privacy controls and shared responsibility would need attention. Its reported score is also the system author’s own measure, rather than an independent benchmark.

Products Worth Considering

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

As an affiliate, we earn on qualifying purchases.

From Session Context to Stored Memory

An AI agent generally responds using information available in its current context window. Meyer says that when a session ends, this context is lost, leaving him to restate decisions and project details in later sessions. He describes existing memory features as often difficult to inspect or specific to individual tools; that is his explanation for building a separate shared store.

The design keeps the memory in ordinary text files and uses an index to direct retrieval. Git records file history, while the scripts provide a consistent route for looking up and adding information. Meyer also describes a triage process that reviews session captures and decides whether to keep, merge or discard them. The article presents this as his implementation and workflow, not as a general test of memory systems.

As of November 21, 2026, Meyer reported 1,278 pages, 1,761 links and 1,363 recalls. He put top-1 recall—the system’s measure of how often its first selected page is the right one—at 97.4%. The published material does not provide a testing protocol, sample size or independent verification for that figure.

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

— Thorsten Meyer, writing on ThorstenMeyerAI.com

How the Recall Score Was Measured

The article does not explain how Meyer calculated the 97.4% top-1 score, what questions or dataset were used, or whether the result was checked by an independent party. The meaning of the reported 1,363 recalls is also not tied to a stated time window. The figures should be understood as metrics Meyer reports for his own system, not as a comparative benchmark.

It is also unclear how the setup handles access controls, concurrent edits or data retention if used by multiple people. Meyer says the system is built for a single owner, and the source does not describe a team deployment, formal security review or comparison with other memory approaches. The article says personal data in its video and frames is protected, but gives no broader account of privacy safeguards.

Further Testing and Team Use

Meyer’s article includes an explainer video showing the system’s views and workflows. It does not announce a product release, public repository, independent evaluation or planned team version. The next clear step for readers seeking to assess the claims would be more detail on the recall test, including its method, dataset and measurement period.

For now, the described system remains a personal implementation. Whether the same file-and-index design can support shared organizational use, with suitable permissions and reliable collaboration, is not established by the material published November 21, 2026.

Key Questions

What is Meyer’s Second Brain?

It is a personal knowledge base that Meyer says he shares with the AI agents he runs. The agents can retrieve information from it and add new material.

How does an agent find information?

According to Meyer, recall.py scores entries in the index first, opens the highest-scoring page and returns the relevant section with a source reference.

What technology does the system use?

Meyer describes Markdown files stored in Git, an index file and scripts for recall and updates. He says it does not use a vector database or cloud service.

Is the 97.4% recall score independently verified?

The source does not provide an independent evaluation or enough detail about the test method. It should be treated as a figure reported by Meyer for his own system.

Can a team use the same setup?

Meyer says the current design is built for one owner, not a team. The article does not establish how it would handle team permissions or concurrent changes.

Source: ThorstenMeyerAI.com

You May Also Like

VigilSAR Benchmark: There Is No Best Model

The VigilSAR Benchmark reveals there is no universally best AI model for defense, emphasizing context-specific rankings based on capability, reliability, and compliance.

The $9 Billion Signature Tax: How DocuSign’s Business Model Survives on One Assumption

A new open source project, DocuSeal, challenges DocuSign’s dominance by offering a free, self-hosted digital signature solution, raising questions about industry reliance on proprietary pricing.

World Model Readiness: Are You Ready for AI That Acts?

Assess your organization’s readiness for AI systems capable of predicting and acting, as the shift from descriptive to action-oriented models accelerates.

Why Does Musk Say Grok Trails Anthropic By Years While His Agentic AI Surges?

A Wccftech headline attributes two claims to Elon Musk, but the available material lacks a transcript, comparison details and growth data.