Vault v2: Capture Works. Search Is the Hard Part.

Today, I sent a note from Claude on my iPad and watched it arrive in my real Obsidian vault. The system saved it, confirmed its location and backed it up through Git.

After the longest PersonalOS build I’ve attempted, that ordinary moment felt significant.

The bigger goal is continuity: one body of knowledge across my phone, iPad, work laptop and Mac mini, accessible through Claude, Codex and eventually other AI interfaces. I want to change devices or models without rebuilding the context of my life.

One part is working

Remote capture is now live in a limited rollout.

I ask Claude to save something. The service queues it, my Mac mini turns it into a note, and I can check whether it is waiting or saved.

We’ve verified the complete path, including retries, backup and stopping and restarting the worker. The initial trial accepts ordinary personal notes and has a temporary three-capture daily cap.

The foundation stays simple: readable Markdown files that I own. AI provides another way to work with them.

The difficult part is finding the right information

I expected a purpose-built search index to make answers faster and better. So far, it hasn’t consistently beaten ordinary file tools.

One lesson stands out: finding the right note isn’t the same as finding the answer.

A search might locate the correct meeting but miss the sentence saying a task was already completed. It might surface an old project summary instead of the later conversation that changed the plan.

And some questions aren’t really searches. “Show me every open opportunity” needs a complete filtered list. “What changed?” needs evidence across time. “What should I pause?” requires judgment.

Treating all of those as “find a few relevant paragraphs” was too simplistic.

We’ve tested approaches that didn’t work

We tried richer excerpts, fewer duplicate results, different ranking methods and a small local model that searches by meaning.

Some improved individual results. None demonstrated enough overall benefit to release.

That matters. It would be easy to show a convincing example and call the problem solved. Instead, we’re measuring answer quality, supporting evidence, speed and how much information the AI has to read.

The benchmarks remain unchanged. Failed experiments stay in the record.

Production taught us another lesson

The test system worked before the deployment process was ready.

The rollout exposed script errors, configuration assumptions and a build rule that quietly skipped releases. Recovery checks caught the problems, but several should have been found earlier.

A working demo and a dependable system are different accomplishments.

We’re improving both: the software itself and the process that installs, verifies and recovers it without constant supervision.

What we’re trying now

I researched how other systems handle similar problems. The useful pattern is to combine methods: structured filters for lists, search for discovery, and direct reading for evidence.

The experiment running now adds a model that takes an existing set of possible passages and puts the most useful ones first. It runs locally on the Mini, with network access blocked during the test.

There’s no result yet. If it helps, it earns another test—not an automatic deployment.

Local testing also doesn’t dictate the finished architecture. I want cloud querying to remain available when the Mini is offline, with a clear indication of how current its information is.

Where this leaves us

Capture works today. Retrieval still has to earn its place. Task integration and the practical setup for daily use can move forward independently.

I’m aiming for a useful version by Monday, without pretending every capability will be finished.

I’ve also become less impressed by how many agents are running or how many tokens they consume. The output is what matters.

Today’s output was a note, saved correctly, from a device that wasn’t sitting at my desk.

The next milestone is an answer I can trust.