Every conversation I have with an AI assistant about this house starts the same way it started the very first time — with nothing. It doesn't remember that we already tried the Tesla integration three different ways, or that the search box only matches display names, or that raising the battery reserve at midnight is too late because the EV automation already fired. I'd re-explain the same hard-won lessons over and over, like Groundhog Day for home automation.
So this weekend, I finally fixed it — properly, not with a workaround.
Not a database. A memory.
Most "AI memory" is really just a search index: you ask a question, it finds documents that sound similar, and hands them back. That's not how remembering actually works. When you recall something, it doesn't arrive alone — it drags related things along with it. Ask about the battery, and you should also get reminded about the EV automation that depends on it, without having to ask that separately.
The system I set up this weekend works closer to that — spreading activation, if you want the technical term. One memory lights up the ones connected to it. It also forgets, gently: old, unused details fade over time the same way they would in an actual memory, while the important stuff — the load-bearing facts about how this house actually works — stays put.
Self-hosted, on purpose
This runs on our own hardware, not a third party's server. That mattered to me more than I expected it to — this is a system that ends up knowing quite a lot about how our home works, and I wanted it living somewhere I actually control, behind a firewall I actually configured, not floating around in someone else's cloud.
It also means every night, while the house is quiet, it "dreams" — reviewing the day's memories, consolidating the ones that turned out to matter, and letting the noise settle out. I like that framing more than I expected to. It's a nice thing to know is happening overnight.
Honest shout-out here: the thinking behind spreading activation, salience decay, and the whole "memory that dreams" model isn't mine — it's the work of the team behind Adja. I just get to run it. If any of this sounds interesting, it's worth a look at what they're building.
What's next
Update: Adja now gives Codex, Claude, and Grok access to the same self-hosted memory, each through its own securely scoped login. That means the useful context follows the work across assistants instead of being trapped in whichever chat happened to discover it.
The next project is giving the rest of the family a proper way in. There's genuine family interest in seeing how the automations are actually built, which feels like exactly the kind of thing worth encouraging. More on that soon.