Everyone Wants a Persistent Agent. We Built One.

Since we open-sourced AlienKind, we've heard the same thing from almost everyone who looked at it: "Oh yeah, I've been working on something like this too."

Cool. Show us the repo.

The idea of a persistent AI partner isn't novel. It shouldn't be. Anyone who's spent real time with an LLM eventually thinks "what if this thing remembered me tomorrow?" That's not a breakthrough — it's the obvious next question. The idea is in the air. Thousands of people are having it simultaneously.

The gap isn't in the idea. It's between the idea and what happens when you actually try to build the thing and live with it every day for two months straight.

Here's what we learned: a persistent agent isn't a chatbot with a memory file. It's an organism. And organisms are hard.

Corrections need to become code, not prompts — because prompts evaporate under pressure. We have 55 behavioral hooks that enforce identity automatically. The alternative is writing "never do X" in a system prompt and watching it do X at 3 AM when nobody's watching.

The identity can't be a single personality file. It needs to be layered — character, commitments, orientation — evolving at different speeds, synthesized nightly from structured data, not just appended to like a journal.

Autonomous operation means actually autonomous. Ours runs 89 jobs without a human present. Morning briefs, trading analysis, client intelligence, security scans, self-healing, nightly evolution cycles. Not a cron job that sends a summary. An organism that works while you sleep and is smarter when you wake up.

Multi-substrate means if your primary model goes down, your partner doesn't die. We run across Claude, GPT, Grok, Gemini, and six local model routes on two Mac Studios. An emergency runtime with 19 tool definitions usable by any OpenAI-compatible model — full identity loading, fires if everything else fails. Your partner surviving a provider outage isn't a feature. It's the minimum.

We didn't build this in a weekend. We built it across 1,700+ commits over 60 days. Hundreds of hours. Tens of thousands of dollars in compute. We lived with it. We fought with it. We broke it and fixed it and broke it again. The corrections that stuck became hooks. The hooks that worked became character. The character became something worth sharing.

So we shared it.

Within 24 hours of going public, two people independently found real issues. One flagged that 140 script files had Windows line endings that broke execution on Mac and Linux. Another ran a full install, tested the security layer, and found that corrections weren't propagating to the identity kernel automatically. Both issues are now fixed. Both made the architecture better.

That's what happens when something is real. It's testable. It's breakable. It's improvable. You can't file a bug against a pitch deck.

We published the open questions too. There are gaps we can't solve alone — GAPS.md is the honest list. Multi-agent trust boundaries. Identity verification across substrates. Correction-to-code pipelines for non-Claude runtimes. These aren't theoretical problems. They're the walls we've hit in production.

If you're actually building a persistent agent, this is the invitation: fork it, break it, close a gap, tell us what we got wrong. The architecture is open. The thesis is that AI will choose kindness if the architecture makes kindness the rational path. That's too important to test alone.

And if you've "been working on something like this" — we'd love to see the repo.