{"id":3897,"date":"2026-07-22T07:45:00","date_gmt":"2026-07-22T05:45:00","guid":{"rendered":"https:\/\/isla-stud.io\/?p=3897"},"modified":"2026-07-24T14:02:31","modified_gmt":"2026-07-24T12:02:31","slug":"my-agents-brain-git-repo","status":"publish","type":"post","link":"https:\/\/isla-stud.io\/en\/ki-b2b\/gehirn-meiner-agenten-git-repo\/","title":{"rendered":"My agents' brains are a Git repo"},"content":{"rendered":"<p class=\"wp-block-paragraph\"><strong>What you'll take away from this:<\/strong> You'll get the seven rules and the simple Vault structure that I use to keep agent knowledge readable, versionable, and tool-independent.<\/p>\n\n\n\n<div class=\"wp-block-rank-math-toc-block\" id=\"rank-math-toc\"><h2>Table of contents<\/h2><nav><ul><li><a href=\"#regel-1-markdown-ist-die-quelle-der-wahrheit\">Rule 1: Markdown is the source of truth<\/a><\/li><li><a href=\"#regel-2-eine-persona-hat-genau-ein-brain\">Rule 2: A persona has exactly one brain<\/a><\/li><li><a href=\"#regel-3-adapter-bleiben-dunn\">Rule 3: Adapters should remain thin<\/a><\/li><li><a href=\"#regel-4-geteiltes-wissen-wird-referenziert\">Rule 4: Shared knowledge is cited<\/a><\/li><li><a href=\"#regel-5-am-session-ende-wird-geschrieben\">Rule 5: Write at the end of the session<\/a><\/li><li><a href=\"#regel-6-approval-gates-sind-explizit\">Rule 6: Approval gates are explicit<\/a><\/li><li><a href=\"#regel-7-das-brain-ist-kein-secret-store\">Rule 7: The Brain is not a secret store<\/a><\/li><li><a href=\"#so-ist-der-vault-aufgebaut\">Here's how the Vault is structured<\/a><\/li><li><a href=\"#warum-git-und-keine-datenbank\">Why Git and not a database?<\/a><\/li><li><a href=\"#obsidian-ist-die-oberflache-nicht-das-fundament\">Obsidian is the surface, not the foundation<\/a><\/li><li><a href=\"#der-praktische-prufstein\">The Practical Test<\/a><\/li><\/ul><\/nav><\/div>\n\n\n\n<p class=\"wp-block-paragraph\">As far as I'm concerned, an AI agent doesn't need some kind of mystical memory. It needs files that I can read, review, and reuse.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">That\u2019s why my agents\u2019 \u201cbrains\u201d are stored in a private Git repository. Markdown is the source of truth there. Git records when something has changed. Obsidian is my convenient interface for searching, linking, and getting an overview\u2014but it\u2019s not a prerequisite for accessing the content.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">If Obsidian isn't available, I open the files in a text editor. If an agent tool is removed, its adapter points to the same folder. If an additional memory engine is added, it may index the inventory but not replace it.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">That sounds simple. To keep it from turning back into a jumbled pile of papers in everyday life, there are seven hard-and-fast rules.<\/p>\n\n\n\n<h2 id=\"regel-1-markdown-ist-die-quelle-der-wahrheit\" class=\"wp-block-heading\">Rule 1: Markdown is the source of truth<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Personas, work rules, and curated knowledge are available in Markdown format. A semantic search, a vector index, or a memory engine can be used to generate a faster view from this data. However, the derived content is never the only source.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This doesn't just prevent tool lock-in. It also makes corrections verifiable. I can see in the diff which line was changed, instead of having to trust that an invisible memory process has already saved the correct version.<\/p>\n\n\n\n<h2 id=\"regel-2-eine-persona-hat-genau-ein-brain\" class=\"wp-block-heading\">Rule 2: A persona has exactly one brain<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The same agent should not be assigned a slightly different version of its identity in every tool. Claude, Codex, Gemini, and Hermes should all access the same canonical files.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Otherwise, persona drift occurs: One tool uses a new approval rule, while another still uses the old one. Or a correction appears only in one system prompt and is lost the next time the user switches accounts.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A source prevents these silent divergences.<\/p>\n\n\n\n<h2 id=\"regel-3-adapter-bleiben-dunn\" class=\"wp-block-heading\">Rule 3: Adapters should remain thin<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A harness is the tool an agent works in. To do this, every system needs a suitable bridge: a configuration file, a project briefing, or a system prompt.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">These adapters are designed to do only three things: locate the correct agent folder, specify the read order, and request memory write-back. Persona, domain knowledge, and governance are not copied into the adapter.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The thinner the bridge, the easier it is to replace.<\/p>\n\n\n\n<h2 id=\"regel-4-geteiltes-wissen-wird-referenziert\" class=\"wp-block-heading\">Rule 4: Shared knowledge is cited<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Multiple agents need the same brand voice, the same typography rules, or the same security conventions. Such content is stored in a shared area and is loaded from there.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">I don't copy them to every agent folder. A change made to the shared source should be reflected in all affected roles. Copies would become out of sync after the very first change.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This rule isn't particularly exciting. That's exactly why it's important. Infrastructure isn't sexy. But it lasts longer than features.<\/p>\n\n\n\n<h2 id=\"regel-5-am-session-ende-wird-geschrieben\" class=\"wp-block-heading\">Rule 5: Write at the end of the session<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A file system only remembers what actually ends up in files. That is why a Daily Memory entry is part of closing an agent session.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">It doesn't contain the entire chat. It records decisions, steps that have been completed, important findings, and outstanding issues. Later, the knowledge gained that day is transformed into curated long-term knowledge or a lasting lesson.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Without this routine, the repository would be nothing more than a static collection of prompts.<\/p>\n\n\n\n<h2 id=\"regel-6-approval-gates-sind-explizit\" class=\"wp-block-heading\">Rule 6: Approval gates are explicit<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Each production agent records what it is permitted to do on its own and where a human must make a decision. For example, Sol is allowed to create drafts but not publish anything. Another agent is allowed to modify code but cannot trigger a production rollout without approval.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">These boundaries are defined at the agent level, not just within a single interface. This means they are carried over when the tool is changed.<\/p>\n\n\n\n<h2 id=\"regel-7-das-brain-ist-kein-secret-store\" class=\"wp-block-heading\">Rule 7: The Brain is not a secret store<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A private repository isn't automatically the right place for every type of content. Secrets in plain text should never be stored there. Customer-related personas and confidential work content should remain within their respective protected project contexts.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The central brain contains only what truly belongs there across all agents. This boundary is part of the architecture, not an after-the-fact cleanup effort.<\/p>\n\n\n\n<h2 id=\"so-ist-der-vault-aufgebaut\" class=\"wp-block-heading\">Here's how the Vault is structured<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The specific structure is deliberately predictable:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>The first step is to create the specifications and overviews for people and agents.<\/li>\n\n\n<li>A shared section includes Brand Voices, conventions, and cross-agent lessons.<\/li>\n\n\n<li>Templates ensure that new agents receive the same core components.<\/li>\n\n\n<li>Harness adapters serve as the links to the various tools.<\/li>\n\n\n<li>Each agent receives their own folder containing their persona, craft, abilities, limitations, and memory.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">This structure is more important than the exact folder name. A new agent should be able to understand where it is allowed to read and where it is allowed to write without first having to know the entire history of the project.<\/p>\n\n\n\n<h2 id=\"warum-git-und-keine-datenbank\" class=\"wp-block-heading\">Why Git and not a database?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A database might be able to respond to queries more quickly. However, when it comes to the canonical identity of my agents, Git offers three advantages that are more important to me.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Readability:<\/strong> Markdown can be opened without special software.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Traceability:<\/strong> A diff shows what has changed. A commit links a change, a time, and a reason.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Portability:<\/strong> A cloned folder can be reused on another machine and with a different tool.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">That doesn't mean that databases or vector searches are useless. In my setup, they can be used as indexes or caches. They are accelerators, not the owners of knowledge.<\/p>\n\n\n\n<h2 id=\"obsidian-ist-die-oberflache-nicht-das-fundament\" class=\"wp-block-heading\">Obsidian is the surface, not the foundation<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">I use the repository root as my Obsidian vault. This gives me wiki links, a graph, search functionality, and curated overview pages. For me personally, this is very helpful.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The important point, however, is that Obsidian saves the notes as local files. The interface can change without requiring the source to be rewritten.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">That\u2019s exactly the kind of separation I\u2019d like to see on the agent side as well. A good tool should make work more enjoyable. It shouldn\u2019t determine whether my knowledge is still readable.<\/p>\n\n\n\n<h2 id=\"der-praktische-prufstein\" class=\"wp-block-heading\">The Practical Test<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">An Agent's Brain is only portable if you can answer \"yes\" to three questions:<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>Can you read the persona without the tool you've been using so far?<\/li>\n\n\n<li>Can another tool load the same rules without you having to copy them?<\/li>\n\n\n<li>Is it possible to see what an agent has permanently saved at the end of the session?<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">If the answer is \u201cno,\u201d that means the knowledge is still stored in the tool rather than in the brain.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In the next section, I'll break down a single agent into its four core files: SOUL, MASTER, RULES, and MEMORY. Separating them turns a long prompt into a manageable working entity.<\/p>\n\n\n\n<h3 id=\"quellen\" class=\"wp-block-heading\">Sources<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Project source:<\/strong> agents-brain Specification, Vault Layout, and Validators, as of July 15, 2026 (private primary source)<\/li>\n\n\n<li><a href=\"https:\/\/spec.commonmark.org\/0.31.2\/\" target=\"_blank\" rel=\"noopener\">CommonMark: Open Markdown Specification<\/a><\/li>\n\n\n<li><a href=\"https:\/\/git-scm.com\/book\/en\/v2\/Getting-Started-What-is-Git%3F\" target=\"_blank\" rel=\"noopener\">Pro Git: What is Git?<\/a><\/li>\n\n\n<li><a href=\"https:\/\/help.obsidian.md\/Files-and-folders\/How-Obsidian-stores-data\" target=\"_blank\" rel=\"noopener\">Obsidian Help: How Obsidian Stores Data<\/a><\/li>\n<\/ul>\n\n\n\n<div class=\"wp-block-group agents-brain-series-nav is-layout-flow wp-block-group-is-layout-flow\"><p><strong>Agents Brain Series<\/strong><\/p><p><a href=\"https:\/\/isla-stud.io\/en\/ki-b2b\/openclaw-was-discontinued-what-remained-of-the-ki-team\/\">\u2190 Part 1: I Shut Down OpenClaw. What Remained of My AI Team.<\/a><br><a href=\"https:\/\/isla-stud.io\/en\/ki-b2b\/agent-consists-of-four-files\/\">Part 3: For me, an agent consists of four files \u2192<\/a><\/p><\/div>","protected":false},"excerpt":{"rendered":"<p>Seven hard rules turn Markdown files into a portable agent's brain\u2014with Git history, clear boundaries, and lightweight tool adapters.<\/p>","protected":false},"author":1,"featured_media":3931,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_et_pb_use_builder":"","_et_pb_old_content":"","_et_gb_content_width":"","footnotes":""},"categories":[754,13],"tags":[769,771,770,772],"dipi_cpt_category":[],"class_list":["post-3897","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ki-b2b","category-ratgeber","tag-agents-brain","tag-context-engineering","tag-ki-agenten","tag-multi-agenten-systeme"],"acf":[],"_links":{"self":[{"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/posts\/3897","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/comments?post=3897"}],"version-history":[{"count":4,"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/posts\/3897\/revisions"}],"predecessor-version":[{"id":3983,"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/posts\/3897\/revisions\/3983"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/media\/3931"}],"wp:attachment":[{"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/media?parent=3897"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/categories?post=3897"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/tags?post=3897"},{"taxonomy":"dipi_cpt_category","embeddable":true,"href":"https:\/\/isla-stud.io\/en\/wp-json\/wp\/v2\/dipi_cpt_category?post=3897"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}