A local, offline, zero-cost layer that lets Claude recall your notes by meaning instead of keywords. Built over several days; finished, fixed, and verified today.
A search engine for your own brain. Your second brain holds roughly 840 notes (meetings, people, projects, decisions). Until now, when Claude needed to recall something, it scanned a catalog of note descriptions and grep'd for exact keywords. This new layer lets Claude search by what a note means, so the right note surfaces even when your question uses completely different words than the note does.
It started with a YouTube video on Claude memory systems. The useful frame from it: any memory system has three jobs.
The old recall path matched words. If a note about support costs said "cost to serve" but you asked about "support efficiency target," a keyword search could miss it entirely. Semantic search matches meaning, so related ideas connect even when the wording differs. That was the gap worth closing, and it is the one piece that makes recall feel like the system actually knows what is in your vault.
Five small pieces, all running locally on your Mac. No cloud, no subscription, no data leaving the machine.
A free program that runs AI models locally on your Mac. Always on in the background, nearly weightless when idle.
The "embedding" model. It reads text and turns it into a list of numbers that represent its meaning. 1.2 GB on disk.
Walks your whole vault nightly and fingerprints every note. Incremental, so it only re-does notes that actually changed (about 2 seconds).
Takes your question, fingerprints it the same way, and returns the closest-matching notes in about 0.3 seconds, each with its file path.
The searchable file of all those fingerprints (about 4,378 chunks), kept in a hidden folder inside your vault so Obsidian ignores it.
Folded into your existing 3:45 AM catalog job. No new automation. If the engine is ever down, it skips quietly and search falls back to grep.
Your vault is personal, so privacy matters. A local model means it is free, works offline, and nothing ever leaves your Mac. Claude stays the reasoner; this model only does the meaning-matching for retrieval. Anthropic has no embeddings endpoint, so "use my Claude subscription for this" was never an option anyway.
The build came over in a handoff doc from a Slack session. The Slack bot could not finish the last step (it is blocked from editing your core config), so I took over, verified everything, and closed it out.
Claude scanned a catalog of note descriptions and grep'd for keywords. If your words did not match the note's words, the note was missed. Recall depended on guessing the right search term.
Claude searches by meaning across the full text of every note. The right note surfaces even with different wording, ranked by relevance, with the source cited. The catalog and grep remain as a safety net.
| Benefit | What it means for you |
|---|---|
| Better recall | Finds the right note by meaning, not just matching keywords. |
| Private | 100% local and offline. Nothing leaves your Mac. $0 forever. |
| Resilient | If the engine is ever down, it silently falls back to the old catalog + grep. It never breaks an answer. |
| Self-maintaining | Re-indexes itself every night, only re-fingerprinting notes that changed. |
No. Here is the honest, measured answer, taken from your machine right now.
Ollama runs quietly in the background, but it holds nothing in memory when you are not searching. When Claude runs a search, the model loads, answers in about a third of a second, and unloads itself after a few idle minutes. Against 36 GB of memory, the footprint is a rounding error. You will not notice it during normal work.
What is installed and running today (the BGE-M3 search model) is tiny and idle-friendly. It does not lag your Mac. The thing you may be worried about, a heavier model, is the optional item below that we did not install.
There is exactly one open item, and it is optional.
A local reasoning model, separate from the search model. It is "insurance" so you could still get AI answers if the cloud ever went down. We deliberately did not install it. If you ever did, it would use roughly 8 to 9 GB of memory only while actively answering a question, then unload. On your 36 GB Mac that is comfortable, but it is a break-glass backup, not a replacement for Claude. Recommendation: skip it for now. You can add it anytime with one command, and it changes nothing until you ask for it.
To be completely clear on the lag question: nothing that touches the heavier model is installed, so the impact on your computer today is zero. Everything described above is the lightweight search layer only.