Six months into using AI heavily with my notes, I caught myself staring at an Obsidian entry I couldn't remember writing.
Not in the "wow, I forgot I wrote that" sense. In the practical sense: I genuinely couldn't tell whether the paragraph in front of me was a thought I'd had, or an AI summary I'd accepted, or a paraphrase I'd nodded along with somewhere between the two.
The skill was intact. I could still write, still think, still synthesize. What had quietly dissolved was the boundary between what I think and what I read and accepted.
This post is about why that happens, why "AI should elevate, not replace your thinking" needs a more practical translation, and the three-layer rebuild — plus an open-source Obsidian plugin — that fixed it for me.
The Failure Mode Isn't Skill Atrophy. It's Authorship Loss.
The mainstream AI critique frames the danger as skill atrophy: lean on the LLM too much, your underlying ability rots. There's something to that, but it wasn't my failure mode.
My failure mode was authorship loss — and authorship loss is sneakier, because it doesn't show up in your ability to perform tasks. It shows up months later, when you're trying to retrieve a position you held and realize you can't reconstruct whose position it actually was.
If you outsource thinking to AI, you can rebuild the skill. If you let AI ghostwrite into the same container as your own thinking, the contamination is permanent. There's no clean way to subtract back.
What Went Wrong: A Single Vault for Everything
My original setup was elegant on paper. One Obsidian vault. AI-generated summaries flowed in (article distillations, meeting notes, reading reflections). My own writing flowed in. Both were tagged, linked, semantic-searched.
The problem: AI output and human authorship have wildly different epistemic statuses, but in plain markdown they look identical. Six months of accumulation later, I had thousands of notes where the answer to "did I think this, or did the AI summarize this?" was effectively unrecoverable.
The "subtle wrongness" wasn't immediate. It crept in via dozens of small moments where I'd skim an AI summary, internally agree, and let it sit in my vault. Later, the agreement got remembered as a position. The position got referenced. The reference got built upon.
The Three-Layer Rebuild
I rebuilt with explicit boundaries:
Layer 1: AI for fast processing.
LLM tools handle ingestion — long article summaries, transcript condensation, dense PDF extraction. High-throughput, low-judgment work. The output stays in a clearly-labeled scratch directory, not the main vault.
Layer 2: A slow reflection layer where I rewrite in my own words.
This is the step everyone skips. Nothing crosses from Layer 1 to my main vault without me rewriting it. Not paraphrasing — rewriting. Different sentence structure, different emphasis, my own framing of why it matters.
The friction of rewriting turned out to be the point. That's where the elevation actually happens. Without that step, AI output passes through me without leaving fingerprints.
Layer 3: Plain markdown as the source of truth.
The main vault is plain .md files in Git. No AI-generated prose lives here unless I've rewritten it. If GitHub disappeared tomorrow, I'd still have my thinking on disk.
The Plugin: AI as Retrieval, Not Generation
The rebuild left one open question: can AI play any role inside the main vault, or must it be exiled entirely?
It can — but only as retrieval, not generation. AI helps me find things I already wrote, without ever generating new content into the vault.
I built Vault Search, an Obsidian plugin that does local semantic search across your notes using embeddings. The use case is the one keyword search keeps failing: I remember thinking about money but my note is titled 預算規劃 (budget planning). Keyword search returns nothing. Semantic search returns the right note immediately.
Three design constraints made it suitable for the three-layer architecture:
1. It only retrieves, never generates.
The plugin surfaces notes you wrote. It will not write new ones. There is no "AI assistant" that synthesizes a response. Synthesis stays your job, with the right material in front of you.
2. It runs locally.
Embeddings via Ollama, 8GB RAM is enough. No cloud API key, no upload, no provider lock-in. If Anthropic / OpenAI / whoever shuts down tomorrow, the index keeps working.
3. AI-assisted indexing, not AI-authored content.
The recommended workflow uses AI to generate 50-100 character semantic descriptions in each note's frontmatter, then indexes those. The descriptions are AI-written but they live in metadata, not in the body. Your prose stays yours.
There's a hot/cold layering on top — recently-edited and well-linked notes get prioritized in search results, so the plugin nudges you toward what you're actively thinking about, not what's been gathering dust for three years.
The Three Rules I Now Follow
The structural translation of "elevate, not replace" comes down to three rules:
Keep AI output in a different container from your own thinking. Different folder, different vault, different file convention — whatever lets you tell at a glance which is which.
Force a rewriting step before anything crosses over. Not paraphrasing — full rewriting. The friction is the elevation.
Constrain AI's role inside your corpus to retrieval. If AI lives in your vault at all, it should help you find what you wrote, never generate new content into the body.
The first two are about preventing contamination. The third is about choosing what role AI gets to play after the boundary is in place.
Closing: Who's Standing at the Shelf
The library metaphor I keep returning to: AI can fetch any book in the library, instantly, in any language. That's a real superpower and I'm not giving it up.
But the person standing in front of the shelves, picking up a book, deciding what it means to them — that role doesn't transfer. The whole point of a personal knowledge base is that someone is doing that standing-at-the-shelf work, and the someone is you.
If the AI quietly takes over the standing-at-the-shelf, you don't have a knowledge base anymore. You have a stranger's library labeled with your name.
Vault Search is open-source on GitHub. The full origin story: why I separated AI from my thinking | building retrieval-only AI for Obsidian.
Top comments (0)