Show HN: Continual Learning with .md

On April 13th, a developer appeared on Hacker News with a project built around an idea so deliberately low-tech it loops back around to interesting: continual learning scaffolded around plain markdown files. Not a new model architecture. Not a fine-tuning regime requiring a PhD and a cloud bill. Just .md files — the same ones developers already use to document their weekend projects — repurposed as structured external memory for an AI system. The premise, as best as can be reconstructed from a single-signal story, is that an AI agent can update its own understanding over time by reading and writing to a set of markdown documents, effectively treating a folder of text files as a persistent, human-readable brain. What makes this worth noting is not novelty in the academic sense. It is that someone built toward a real problem — AI systems forgetting everything between sessions — using the simplest possible substrate. No vector databases. No embeddings pipeline. No infrastructure you need to explain to anyone. Just files a developer can open, read, edit, and understand without tooling. The vibe coding sensibility is obvious here: build the thing that works before building the thing that scales.

If confirmed as a working system, here is what this means. The path to persistent AI agents just got significantly more accessible — not for enterprise deployments, but for the individual developer who wants an assistant that remembers context across projects without standing up a database. The second-order effect is more interesting: if markdown proves sufficient as a memory layer, it applies pressure to the assumption that continual learning requires specialised infrastructure. Developers already living inside markdown-heavy workflows — documentation, note-taking, second-brain systems like Obsidian — would find integration almost frictionless. That is a large population. Toolmakers building on top of language models would face a genuine question about whether they are overcomplicating the memory problem, and investors backing infrastructure-heavy approaches to agent persistence would be wise to at least take the question seriously.

Watch for whether the Hacker News thread surfaces concrete implementation details or working repositories that others can fork and test — and whether anyone outside the original post begins reporting results from trying it themselves.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →