Search for semantic duplicates every time you create a new note
Run semantic similarity searches against your existing notes when creating new notes to detect conceptual duplication hidden behind different vocabulary, treating AI-surfaced matches as candidates for potential abstraction or cross-linking.
Why This Is a Rule
Knowledge bases accumulate conceptual duplicates that keyword search can't detect. You write a note about "cognitive load" and don't realize you already have notes about "mental bandwidth," "working memory limits," and "attention budget" — different vocabulary for overlapping ideas. Over time, these hidden duplicates fragment your thinking: related insights live in disconnected notes that never cross-pollinate.
Semantic similarity search solves this because it matches on meaning, not keywords. When you create a new note, a similarity search against your existing base surfaces notes that express related ideas regardless of vocabulary. These matches are candidates for three actions: merging (the notes say the same thing), abstracting (the notes are instances of a higher-level pattern), or cross-linking (the notes are genuinely distinct but related).
When This Fires
- Every time you create a new permanent note in your knowledge base
- When you notice you're writing something that "feels familiar" but you can't find the original
- During periodic knowledge base maintenance to detect accumulated duplicates
- When migrating notes from one system to another and deduplication matters
Common Failure Mode
Treating semantic matches as false positives because the vocabulary is different. "That note is about cognitive load, mine is about mental bandwidth — they're different." Often they're not. The different vocabulary masks conceptual overlap that, once recognized, produces a more powerful abstraction than either note alone. Dismiss matches slowly, not quickly.
The Protocol
After writing a new note: (1) Run a semantic similarity search against your existing notes (tools: Obsidian Smart Connections, Mem, custom embedding search). (2) Review the top 5 matches. (3) For each match above ~0.8 similarity, decide: merge (same idea, different words), abstract (instances of a pattern — create a parent note), or link (distinct but related). (4) If no matches, the note is genuinely novel — proceed.