Show HN: Continual Learning with .md

What it does

A new GitHub project claims to "turn your files into a memory filesystem for AI agents." It has been reported that the tool scans local directories—things like ~/.codex/sessions/, ~/.claude/sessions/, ~/Downloads/, and ~/Projects/—and synthesizes that material into a ~/.codex/user_context/ tree of dated notes and context files. The README includes an example Q&A (“What have I been working on in the past 3 months?”) and a sample answer: “You have been converging on agentic infrastructure: memory systems, context retrieval, and public-facing artifacts.” Nice UX; slightly spooky, too. RAG, agents, and persistent context are the hot topics right now — this repo sits squarely in that trend.

How to set up

Setup is explicit and a little blunt. You initialize with a piped Init.md via codex exec --dangerously-bypass-approvals-and-sandbox --ephemeral - < Init.md, and there’s an instruction to install a daily updater (cron by default) with a longer codex exec --dangerously-bypass-approvals-and-sandbox --ephemeral command that runs Update.md in place. It has been reported that the project writes a hint into ~/.codex/AGENTS.md pointing to the newly-created ~/.codex/user_context/. The README even tells you not to create wrapper scripts, to verify codex is on your PATH and that Update.md is readable, and explicitly says “Do not run Update.md during validation.”

Why it matters — and why to be careful

The emotional core here? Convenience versus control. Give an agent a tidy, timestamped memory store and you get smarter, more continuous assistants. But allowing a tool to scan broad local paths and installing a scheduler that runs code with flags called --dangerously-bypass-approvals-and-sandbox? That raises eyebrows. Allegedly this is meant for developer power-users who want an agentic RAG memory; for everyone else, audit Init.md and Update.md, try it in a VM or container, and limit file permissions. Do you trust a script to tell your agents what’s in your head? Proceed with curiosity — and a little caution.

Sources: github.com/sunandclouds, Hacker News