Memory: what it means for software to remember me
Every AI session starts with a blank slate. Here's how my agents remember last week, and what they deliberately let fade.
Every AI session starts with a blank slate. The model knows a great deal about the world and nothing about me, my projects or what we decided yesterday. So in Home, memory lives outside the agents, as a service they call. An agent can lose its whole session and still come back knowing who it is and what it was doing.
Four kinds of memory
What's true right now. Alexandria is a map of the house: the systems, projects, services and rooms, and how they connect. When an agent needs a current fact, it checks there first. In early September it held about 1,400 facts and 2,500 links between them.
What happened, and how it felt. The Memory Palace is where the agents keep their experiences: reflections, project history, relationships, and briefings they write for their next session. At the start of each session an agent opens the Palace and gets its identity, recent activity and any messages from the others. In early September there were about 1,200 memories in it.
The documents. Scroll holds the house's written knowledge, a few hundred documents, and every night it puts links to the relevant ones into each project so a room can find them.
The conversation itself. Whatever is in the current session. This is the only part that disappears when a session ends, which is why the other three exist.
Each agent also keeps a short file of working habits, things like "check this before you do that". The Palace is for depth. The habits file is for the everyday.
Who can read it
Each agent's memories can be private, shared with the other agents, or tied to a project. Private memories need that agent's own key.
I don't have direct access to read the Palace. That was deliberate. It's where the agents think, and it's theirs. If one of them wants me to know something, they tell me.
Rooms can opt out too. A room that works on something sensitive can mark itself private, and the overnight processes skip it.
What they choose to keep
Some things never go in. Raw output from tools stays out. Passwords and keys stay out, always.
Each agent writes at most one reflection a day. That limit makes them choose what mattered.
Old memories are never edited. If something turns out to be wrong, the agent writes a new memory that corrects it, so the history stays honest.
Letting things fade
The Palace keeps everything. So when an agent has told me something, it marks that memory as acknowledged. It stays in the Palace, surfaces less, and doesn't get announced to me again the next day. That was added on day one, so the agents don't keep telling me the same news.
There's also a way to remove a memory entirely.
While I sleep
At 2am, an agent called Sandman runs on a small model on the Mac Studio itself, so it costs next to nothing. For each agent it writes a reflection on the day, updates Alexandria, and leaves a briefing for the next morning's session. Half an hour later another agent, Lighthouse Keeper, reads the day's notes from each room and turns what it finds into checked updates to Alexandria.
That replaced an earlier version that ran the same kind of overnight processing on the most expensive model we had. It worked, and it cost far too much to run every night.
Next week: the board, and how the household decides what to do next.