May 14, 2026
Git happens, drama follows
Best "Brain" for Agents Is Just Versioned Folders of Markdown Files
AI memory craze ends with... folders of notes, and the comments are losing it
TLDR: The headline-grabbing idea is that AI agents may work best with simple versioned notes instead of expensive custom memory systems. Commenters were split between calling it obvious common sense and mocking the claim as clickbait that still depends on extra search tools.
The big plot twist in AI land? After spending two years building fancy systems so chatbots could "remember" things, engineers are now crawling back to the digital equivalent of a labeled filing cabinet: markdown notes in folders, saved with version history. The article points to GBrain, an open system from Y Combinator boss Garry Tan, plus a Hacker News side quest called DiffMem, as proof that the boring answer may actually work better in real life. Why? Because companies got tired of sky-high bills, broken memory, and bots that still acted like they had the attention span of a goldfish.
But the real show was in the comments, where readers instantly split into camps. One group basically yelled, "Well... duh," with one commenter saying this should be obvious to anyone who has used a coding agent for more than a few hours. Another group was not buying the oversimplified headline for a second, accusing it of glossing over the fact that these systems still rely on search layers and extra structure. The sharpest jab? "This isn't even well-slopped slop," which is so rude it almost loops back around to poetry.
The messiest drama centered on whether Git — the tool programmers use to track file changes — is really replacing fancy AI search tools at all, or whether that's just hype with better branding. Critics called the title clickbait, skeptics asked if this would waste tons of computing power, and pragmatists shrugged: put your notes in Git, add decent search, move on. In other words, the crowd's verdict was classic internet: half revolution, half eye-roll, fully entertained.
Key Points
- •The article argues that persistent AI agent memory problems are primarily organizational knowledge issues rather than simple context-window limitations.
- •It presents GBrain, open-sourced by Garry Tan in April 2026, as a markdown-and-Git-based system with PostgreSQL and pgvector for hybrid search.
- •GBrain is described as using a page structure with an editable summary and an append-only timeline to preserve evidence and support multi-agent collaboration.
- •The system includes an automated nightly “dream cycle” and extracts typed knowledge-graph links from page writes without LLM calls.
- •The article cites BrainBench results showing GBrain achieved P@5 of 49.1% and R@5 of 97.9%, and references DiffMem as an earlier Git-and-markdown proof of concept using BM25 search.