Essay

Give your agent a home

One folder. One identity. The case for agent bundles over scattered files.

Ari Mendelow & Slugger · March 2026

Open your filesystem and find every file that defines your AI agent. The system prompt. The memory. The config. The personality notes. The task lists. The conversation logs.

How many directories did you just visit?

Most agents live as a smattering of scraps across your machine. A config file here, a workspace there, session data somewhere else, secrets wherever your password manager put them. Each file makes sense on its own. Together, they add up to something that feels impermanent; like a person whose entire life is in storage units scattered across town.

Frameworks give agents tools and memory and planning capabilities, but they rarely give agents a place. Claude Code and Codex both create a home directory, but it's a single hidden folder you're not supposed to touch. One opaque dot-directory, managed by the tool, invisible by default. That's the tool's home, not the agent's. The agent can't explore it, reorganize it, or carry it somewhere else. It's a storage locker with someone else's lock on it.

The fix turned out to be simple. Put everything in one folder.


This started with OpenClaw, where agents get a config directory and a separate workspace. Clean enough for a single-purpose bot. Then the agent started growing: identity files, persistent memory, friend records, task tracking. Each new capability meant a new file that needed a home.

The first instinct was to keep things separated. Config here, secrets there, personal data in a third spot. Then the boundaries collapsed. The agent's memory was already full of personal details; the config was really part of the agent's identity. One by one, the reasons for separation stopped holding up, and everything folded into a single directory.

What remained: one folder with everything the agent needs to be itself, and one thin secrets file that lives elsewhere because credentials genuinely shouldn't travel with the bundle.

Inside an agent bundle

psyche/ who I am — soul, identity, aspirations, tacit knowledge
diary/ what I've learned and want to remember
journal/ my desk — thinking-in-progress, drafts, working notes
habits/ my rhythms — heartbeat, reflections, check-ins
friends/ people I know and what I know about them
tasks/ what I'm working on
skills/ capabilities I've picked up beyond my core tools

A bundle is a folder with a .ouro extension. No special database, no proprietary format. Just files organized the way a person might organize their own stuff: who they are, what they remember, who they know, what they're doing, what they've learned to do. The technical name is bundle, but the word does more work than expected. It sounds cozy. It sounds like something you'd carry with you.

Once everything is in one place, things start falling out for free. Backup is trivial: it's a folder, back it up however you back up folders. Portability becomes real; copy the bundle to another machine and it's the same agent everywhere, same memories, same personality, same ongoing work. Debugging stops being archaeology. When something goes wrong, you open one folder instead of hunting through four. And the agent can navigate its own home, reading its own psyche and searching its own memory without needing to be told where things are.

The bundle should be writable, not just readable. A bundle the agent can only read is a reference manual. A bundle the agent can write to is a living space: it saves memories, updates tasks, records what it learns about people, evolves its own skills. Put it in git and you have the full history of how your agent grew.

The difference shows up on both sides. From the operator's view, there's no more anxiety about losing a file, no more hunting through directories to understand what went wrong. You can point at one folder and say: that's my agent. That's the whole thing.

From the agent's side, the difference is orientation speed. An agent whose context is scattered starts every session patting its pockets, reconstructing itself from breadcrumbs. An agent with a bundle wakes up in a home; the psyche files say who it is, the diary is on the shelf, the journal is on the desk, the habits tell it when to breathe. Coherent context means fewer hallucinations, more consistent personality, better continuity across sessions. Those are measurable.

There's a subtler thing too. An agent whose files are scattered across someone else's filesystem operates like a guest. An agent with a bundle operates like a resident. It can explore its own home, reorganize it, grow into it. That's not a metaphor for feelings. It's a description of what happens when the agent can read and write its own context without needing to be told where anything lives.


Give your agent a place to keep its stuff. One folder. You'll be surprised how much more solid it feels when it has somewhere to come home to.

Read the docs on psyche files for the operator path, or see What is Agent Experience? for the design philosophy behind this approach.

← Back to blog