Why Ten Years of Technical Notes Remain Useless to AI—and How to Fix It

Problem

When prompting an LLM to generate a dialogue‑agent architecture, the model produced a plausible design but lacked any reference to the author’s prior experiments (vLLM prefix‑cache optimization, grammar‑guided generation results, multi‑model separation review). The model could not retrieve any of the thousands of notes stored in Yuque.

Implication

In the AI era, knowledge that cannot be linked, searched, or read by an agent loses value.

Migration to Obsidian

Obsidian stores notes as local Markdown files, enabling direct file‑system access by scripts and LLMs. Its bidirectional linking surface suggested relevant older notes while drafting an architecture, turning isolated documents into a graph of context.

Four‑layer knowledge asset model

Record – capture research, experiments, decisions, retrospectives in detail.

Link – create explicit relationships (why a choice was made, why alternatives were rejected, which prior judgments influence the current outcome).

Call – keep the knowledge in plain Markdown with clear structure so that AI agents can read and process it.

Close‑loop – treat notes as living artifacts: a research note can generate a solution, which becomes a Codex task, produces code skeletons and validation scripts, and the execution results are written back to the note for future decisions.

Importer tool

The open‑source importer https://github.com/ithzhang/yuque-obsidian-importer reads Yuque’s .lakebook export, parses the directory hierarchy and front‑matter, converts lake HTML to standard Markdown, downloads CDN images locally, and preserves the original ordering.

Backup strategy

Real‑time sync of the Obsidian vault to a cloud folder (iCloud or Baidu Netdisk) for multi‑device access.

Version control via the Obsidian Git plugin: periodic git commit and git push to a private repository.

Cold backup: encrypted archive uploaded to object storage for offline safety.

Outcome

After migration, notes can be linked, searched, read by AI, and incorporated into agent workflows, turning ten years of experience into actionable context rather than isolated files.