Exploring Shared Memory For AI Agents
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Exploring Shared Memory For AI Agents 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 describes a personal knowledge base that lets his AI agents share long-term memory through Markdown files and a single index. As of Nov. 21, 2026, he reports 1,278 pages and a 97.4% top-one recall score; the setup is designed for one owner, and its results have not been independently verified.

Thorsten Meyer has described a shared long-term memory system for the AI agents he runs, using Markdown files and a single index instead of a vector database or cloud service. In a report published Nov. 21, 2026, he said the system contained 1,278 pages and had a 97.4% top-one recall score—figures he presents as his own system’s measurements, not an independently audited benchmark.

Meyer calls the setup his Second Brain: a personal knowledge base that both he and his agents can read and update. The goal is to avoid repeatedly supplying agents with project details, past decisions and failed approaches at the start of each new session. He says any agent can look up information in the shared collection and add new material to it.

The memory is made of plain Markdown files kept in Git, with each page holding a cluster of related facts. A file named INDEX.md lists each memory on one line. When an agent needs information, a script called recall.py scores those index lines, opens the highest-ranked page and returns the section it identifies as relevant, along with the page source. Meyer says this process is deterministic and does not open every page during a search.

New information is added through remember.py, which writes the page, its index entry and a log entry together. Meyer says the single-step process is meant to prevent the index and stored notes from falling out of sync. He also describes nine views for monitoring or organizing the system, including a change log, a page-and-link explorer, a triage process for deciding what to keep, and a console for tracking services and machines.

At a glance
reportWhen: Published Nov. 21, 2026; figures are re…
The developmentThorsten Meyer published a report on Nov. 21, 2026, describing a Git-backed memory system shared by his AI agents and reporting its size and recall performance.

A Memory Agents Can Share and Inspect

A shared store can give multiple agents access to the same project history rather than leaving each one dependent on its own session or tool-specific memory. If the approach works as Meyer describes, a user can correct a stored fact once and make that correction available across agents. This could reduce repeated setup and make answers easier to check when the system returns a source page.

The design also makes a trade-off visible. Readable files and Git history offer direct access to the records and their changes, while the system depends on keeping pages, index entries and links organized. Meyer characterizes the setup as built for one owner, not a team; the report does not establish how it would perform with shared permissions, competing edits or larger organizational workflows.

The reported 97.4% top-one score suggests the first page selected by the system is often the one Meyer considers correct, according to his own measure. But the report does not provide an evaluation protocol or independent test results, so the number should be read as a system-specific claim rather than a general measure of agent memory quality.

Samsung 65-Inch 4K OLED Smart TV with Dolby Atmos, 120Hz Refresh Rate, and Amber Protection Bundle

Samsung 65-Inch 4K OLED Smart TV with Dolby Atmos, 120Hz Refresh Rate, and Amber Protection Bundle

KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, Almond Cream

KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, Almond Cream

As an affiliate, we earn on qualifying purchases.

From Session Context to Persistent Files

AI agents typically work from information available in their current context window. Meyer says that once a session ends, that session’s working context is no longer available to the agent, which can leave users repeating background information. Some products offer memory features, but he says he wanted one store he could inspect and use across the agents he runs.

His system’s basic structure is deliberately simple: a folder of text files, an index and scripts for retrieving and adding information. Git provides a history of changes, while the index is used as the first stage of recall. Meyer says the system had 1,278 pages and 1,761 links on Nov. 21, 2026. He also reports 1,363 recalls, though his article does not specify the period covered by that count.

The article describes a workflow for turning session captures into lasting memory: a triage step decides whether material should be kept, merged into an existing page or discarded. Other views summarize recent activity and frequently used pages. These are components of Meyer’s implementation; the report does not say they are part of a commercial product or independently evaluated service.

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

— Thorsten Meyer, describing his system

How the Recall Score Was Measured

The report does not explain how Meyer calculated the 97.4% top-one recall score, how many questions were tested, who judged whether a selected page was correct, or whether the test included difficult or ambiguous queries. The result is attributed to Meyer and should not be treated as an independently verified benchmark.

It is also unclear what time window the reported 1,363 recalls covers, how much effort the system takes to maintain as it grows, or how it would handle multiple users editing the same knowledge base. The article describes a personal setup and does not provide comparative results against other memory systems. Its claims about avoiding index drift describe the normal write process, not a guarantee that errors or incorrect stored facts cannot occur.

Maintenance and Team Use Remain Open

Meyer’s published account sets out the current design and its reported figures but does not announce a product launch, release date or planned external evaluation. The next useful evidence would be a clearer account of the recall test, including its test set and scoring method, alongside results over a stated period.

For readers considering a similar setup, the report points to practical questions rather than a ready-made team solution: how to review captured information, correct mistakes, manage access and preserve the index as the collection expands. Meyer’s article says the system is intended for his own agents, so whether its structure can support teams remains unresolved.

Key Questions

What is Meyer’s Second Brain?

It is a personal knowledge base that Meyer says he and his AI agents share. It stores related facts in Markdown files and uses an index to help agents find relevant pages.

How do agents search the memory?

According to Meyer, recall.py scores lines in the index, opens the highest-ranked page and returns the relevant section with its source. The system does not search every page directly for each recall.

What does the 97.4% figure mean?

Meyer calls it the system’s top-one recall score: how often the first page selected is the one he considers right. His report does not provide enough testing details to independently assess the figure.

Is this designed for teams?

Meyer describes it as a setup for one owner, not a team. The report does not explain how shared permissions or simultaneous edits would be handled.

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

An Honest Account Of Taking An AI Idea To Production

Thorsten Meyer details how an AI-generated video engine was reviewed, tested and revised, and why completion alone did not prove quality.

Grok 4.6: The Frontier Is Now A Price War

Grok 4.6, released by SpaceXAI, shows modest AI improvements but significantly shifts the market with unchanged pricing amid rising competitors.

OpenStreetMap Meets Quest-Based Mapping In StreetComplete

An IdeaNavigator AI brief proposes testing a role-filtered monitor using StreetComplete as an example. It reports no new StreetComplete announcement.

OLED Gaming Monitors: Burn‑In Myths vs Reality

For those curious about OLED gaming monitors, uncover the truth behind burn-in myths and learn how modern technology protects your display from damage.