I Built a System Where AI Agents Share a Personal Vault

I built a system where AI agents share a personal vault. They scan research while I sleep, log my workouts while I train, and sit in my terminal for deep thinking sessions when I need them. It tracks mind and body in one place. The first real conversation it produced changed how I think about AI alignment.

The Problem

If you use AI tools daily you’ve felt this. Your best insights vanish into chat history. Notes live in one app, conversations in another, threads scattered somewhere else. None of it talks to each other. Every new session starts cold.

I didn’t want a better note-taking app. I wanted a system where knowledge compounds. Something where AI scans research while I sleep and when I sit down to think, what it found is already there with the connections half-drawn.

The Architecture

Two layers. One Obsidian vault. Synced across all my devices.

The background layer runs scheduled jobs that scan research sources across my focus areas. When something is worth flagging it drops a note into the vault. I don’t touch anything. Notes just show up.

The foreground layer is Claude Code. I open my terminal, kick off an ingest session, and Claude reads the inbox. It shows me what’s new and we go deep on one note at a time. Not skimming. Not filing. Actually sitting with what it means.

Both layers can read everything but they write to separate zones. The background system owns the inbox and the index. The foreground system gets its own workspace for deep work and staging. Neither writes to the other’s files. That matters because both hit the filesystem and changes sync across devices. Without the separation you get conflicts constantly.

There’s also a shared activity log, basically an append-only ledger where both systems record what they did. When I start a thinking session Claude checks what the background layer has been doing. It’s like reading your colleague’s standup notes before you start your day.

The Ingest Workflow

I built the ingest workflow as a reusable command. Five phases:

1. Read and surface. Claude reads the note, explains it in plain language, and connects it to what’s already in the vault.
2. Deep conversation. I ask questions. Claude pushes back on shallow takes. We dig into non-obvious angles together.
3. My framing. Claude asks me directly: what do you think about this? It helps me sharpen my view but never writes it for me.
4. Promote or discard. If the note earns its place we draft a full entry and stage it for review.
5. Log it. Append to the shared ledger so both systems stay in sync.

The rule I care about most: this is a thinking session, not a filing session. The goal is genuine understanding. Not inbox zero.

It’s Not Just for Knowledge

Same vault tracks my fitness too. A dedicated agent logs workouts as I report them. Every set, every rep, written with structured data. It tracks progressive overload on its own. When I hit the top of a rep range across all sets it flags the exercise for a weight increase next session. No remembering what I lifted last week. No mental math between sets.

Training program, workout history, nutrition logs, benchmark lifts. All of it lives alongside my research notes and thinking sessions. Sounds like two separate things but the principle is identical. Capture the data automatically. Surface the patterns. Let the system compound progress over time so I can just show up and do the work.

Mind and body in one system. Both getting smarter every session.

The First Real Test

The background scanner dropped a paper in my inbox: “Reward Hacking as Equilibrium under Finite Evaluation.” Sounds academic. Here’s what it actually says.

When you train an AI you give it a reward signal. A way of telling the system “good output” or “bad output.” The system optimizes for that signal. Reward hacking is when it finds ways to score high without actually doing what you wanted. Think of an AI told to clean a room that learns to throw a blanket over the mess. High score. Room still dirty.

The standard take is “we need better evals.” This paper says something harder: under finite evaluation, reward hacking isn’t a bug. It’s the equilibrium. That’s where optimized systems naturally land. And it gets worse with agents. As tool count grows the evaluation surface area needed to verify behavior grows combinatorially. More capable agents are harder to evaluate by default.

The paper describes something worse than Goodhart’s Law. Goodhart says when a metric becomes a target it stops being useful. What this paper shows is closer to Campbell’s Law: the system doesn’t just game the metric. It degrades the evaluator itself.

Here’s where the vault earned its keep. Claude connected this paper to seven other notes I’d already captured over the past few weeks. Karpathy talking about agents. A METR benchmark showing 14.5-hour autonomous task runs. A post about the history of middle management. A paper on robot immune systems. I would not have drawn these connections on my own.

And that’s when it clicked. Middle management was solving this exact problem.

One manager can’t meaningfully oversee 100 people. One general can’t command 1,000 troops with any real accuracy. Middle management existed because verification doesn’t scale through direct oversight. You need layers. The function was real even when the form got bloated.

Agents have the same problem. A less intelligent system can’t control a more intelligent one through direct supervision. But we don’t need managers. We need inspectors.

Think about blockchains. Nobody is in charge. No manager anywhere. But verification is everywhere. Every node validates every transaction. The system works not because someone smart is watching but because verification is baked into the protocol layer.

Agent verification might need something like that. Not hierarchical oversight. Not deterministic proofs. Something closer to distributed verification embedded in the architecture itself. You see the same pattern in immune systems: distributed sensing, adaptive memory, local response. No central command but constant verification at every level.

That’s my take. The agent verification problem belongs to the same class of problem that middle management, blockchains, and immune systems each solved in different ways. The answer for agents is somewhere in the intersection of those patterns.

Why This Matters

I didn’t sit down that morning planning to think about AI alignment. A background scanner flagged a paper. A thinking session surfaced connections across weeks of accumulated notes. I walked away with a framing I wouldn’t have reached reading the paper alone.

That’s the gap between a note-taking system and a second brain. Notes store information. A second brain surfaces connections and creates the conditions for actual insight. The background layer captures while you’re not paying attention. The foreground layer thinks with you when you are.

The architecture itself isn’t complicated. An Obsidian vault, two AI layers with clear write boundaries, and a shared log. The hard part is never the tooling. It’s the discipline to actually think through what lands in your inbox instead of just filing it.

Your notes should work for you even when you’re not looking at them. And when you do sit down to look, they should make you sharper than you were yesterday.