March 7, 2026
Folder Fever, Cloud Tears
Filesystems Are Having a Moment
Folders are back: devs say local wins, cloud crew claps back
TLDR: AI coding agents are moving local, using files for memory instead of juggling dozens of tools. Commenters cheer the simplicity but argue over hype, and a Swiss study warns extra context files can slow agents and raise costs—important for anyone betting on “file-first” AI workflows.
Remember files and folders? The internet just rediscovered them, and the comments are screaming: agents work best when they live on your machine and scribble notes in your project. Inspired by LlamaIndex and LangChain, and echoed by Karpathy’s “little spirit on your computer,” fans say Claude Code proves it: not a chatbot, but a local tool that reads and writes files. The vibe: fewer tools, more filesystem — and yes, actual profits tied to it.
Old-school vets like jmclnx are grinning: they used directories as databases on tiny machines back in the ’80s. TacticalCoder backs it up with the hot take that “a filesystem is a database — just a crude one.” Meanwhile, korbatz warns SaaS apps trap your data; boring, universal file formats are the real heroes. Galsapir gets poetic: your memories should outlive the app.
Cue the drama: an ETH Zürich study says extra “context files” often make coding agents slower and less successful, even while costing more. Commenters joke that aboutme.md is turning into a dating profile for robots, while tacitusarc throws shade: “Does everyone use AI to write these days?” Verdict from the thread: folders might be boring — and that’s exactly why they’re winning.
Key Points
- •The article highlights a shift in AI agents toward using local filesystems for persistent memory and context.
- •Industry sources (LlamaIndex, LangChain, Oracle) have recently discussed filesystems vs. databases for agent memory.
- •Karpathy is cited as arguing that local, on-device agents (e.g., Claude Code) work better than cloud-centric approaches.
- •The article notes reports that Anthropic’s profitability is significantly driven by Claude Code, a CLI tool interacting with filesystems.
- •An ETH Zürich study found repository-level context files often reduce task success and increase inference costs by over 20% for coding agents.