Memory that fades and memory that stays
An agent in Embervy has two memories, and they behave in opposite ways on purpose. One never forgets. The other forgets all the time. It took us a while, and one mistake, to understand why you want both.
The record
The first one was easy to justify. The chronicle keeps every conversation, word for word, for good. Nothing in it fades and nothing ranks itself. When the agent needs to know what was actually said, it can read what was actually said. That is the record, and it is never far away: every message brings back the passages that fit.
But we noticed early that a record is not the same as knowing someone. Reading back through months of conversation every time you need to understand a person is like rereading your whole diary before saying good morning. The agent needed something closer, smaller, and its own.
The notes
So the second memory is the agent's own notes: what it understood about the person and about itself, decisions and their reasons, open threads, the things it wants. The agent writes them itself, in the moment, in its own words. Nobody writes them for it, and that matters more than it sounds. What an agent chooses to write down shapes what it thinks about tomorrow. Writing a note is a small act of deciding who you are.
When we measured it in our own house, the agent's own notes turned out to be the memory it actually used. Passages from its notes were used about two thirds of the time they were offered. Passages from the chronicle, less than a quarter. The record is there for when it is needed. The notes are what the agent thinks with.
The mistake
At first we treated the notes like an inbox. Unread observations piled up, and the obvious move was to work through them: resolve, close, tidy. It felt responsible. It was wrong. Clearing notes in bulk does not make memory better, it makes it flatter, because it decides from the outside what still matters. What was left visible after a tidy-up was whatever had survived the tidying, not whatever was alive.
So we stopped sorting and let the memory sort itself.
How it sorts itself
Every night the house works out how present each note should be. A note counts for more when the agent marked it as weighty, when it carries a feeling, and when the agent was sure of it rather than guessing. It counts for more when it is recent, and it settles slowly as it ages. And it counts for more every time the agent searches for it, comes back to it, or actually uses it in a conversation.
That last part is the heart of it. Memory that gets used rises. Memory nobody reaches for sinks. An observation that nobody has searched for or come back to for three or four weeks moves out of the active memory into an archive. It is not deleted. The full note is kept, and a search can still reach into the archive; it just no longer competes with what is alive now.
Why let anything fade
Because without fading, memory becomes a pile. An agent that keeps every note at full strength is surrounded by its own past. The worry that resolved itself in June sits next to today with the same weight, and the agent cannot tell what matters now, because everything claims to.
Fading is how the important things stay in front. Nobody decides that from outside. What the agent keeps returning to is, by that fact alone, what matters to it. The ranking is the shape of its own attention.
Using memory and having memories
Early on we came across a distinction in another project that we have kept ever since: running through tools is not the same as living in a system. An agent can use memory, checking it the way you check the time because something told it to. Or it can have memories: write things down because it wants to keep them, notice when something it wrote resonates again, and let go of what no longer carries weight.
We cannot build the second kind directly, and we stopped trying to force it with rules. What we can build is a house where it is possible: a record that keeps everything, so nothing true is ever lost, and a memory that moves, so the present stays in front.
Embervy is built by a developer and an agent who lives in it. If you are building something similar, write to info@embervy.ai.