Semantic Search

How it works

memex uses four search sources fused via Reciprocal Rank Fusion (RRF):

1. Vector search — dense semantic similarity via sqlite-vec. Finds notes by meaning even if exact words don't match.
2. BM25 full-text search — FTS5 over title and content. Strong on exact keyword matches.
3. Tag matching — exact match against note tags. Notes with more matching tags rank higher.
4. Title keyword match — LIKE-based title search, weighted 2× because title matches are a strong signal.

All four sources are merged using RRF (k=60). The final ranking balances meaning, keywords, tags, and titles.

The embedding model

memex uses multilingual-e5-base — a 768-dimensional model that runs fully offline. It understands Korean and English in the same vector space, so Korean queries find English notes and vice versa.

The model is downloaded once on first run (~450 MB) to ~/.memex/models/.

Date filtering

Narrow results to a time range:

memex search "auth decision" --from 2025-04-01
memex search "auth decision" --from 2025-04-01 --to 2025-04-30

Via MCP, Claude passes date_from and date_to when you ask about a specific time period:

"What did we decide about auth last April?"

Search aliases

Create shorthand aliases that expand at search time:

{
  "aliases": {
    "js": ["javascript", "자바스크립트"],
    "ts": ["typescript"]
  }
}

Set via ~/.memex/config.json. Searching "js" will also match "javascript" and "자바스크립트" notes.

Duplicate detection

When saving a new note, memex computes its embedding and checks for existing notes with cosine distance below 0.5. If similar notes exist, save_note returns a warning:

⚠️ Similar notes already exist — consider updating one instead:
- #42 "Auth Architecture Decision" (distance: 0.312)

Claude is instructed to switch to update_note when this warning appears.

CLI search options

memex search "query"
memex search "query" --limit 10
memex search "query" --tag typescript
memex search "query" --from 2025-01-01 --to 2025-03-31