Introducing AEON
We built a model
that does not
forget.
Every system writes a record longer than anyone can read. Machines, networks, grids,
markets, patients. The history piles up faster than it can be held, so the decision gets
made off the last few minutes instead of the whole record.
AEON does not choose between the two. It carries the record forward as state, compressed
and continuous, while attending to the present window in full detail. What is happening now
is read against everything that came before it, without paying to re-read the past. A new
model, seven years from scratch, small enough to run on hardware you already own.
The core architecture
AEON
A hybrid recurrent transformer, built from scratch over seven years. Not a wrapper, not a fine tune of someone else's weights. It is the common intelligence layer beneath every Horizon product.
AEON holds operational history across the full sequence rather than a fixed window, at a footprint small enough to run on hardware you own. An excursion in March is still there in October, so a failure is not only detected, it is explained by the history that produced it. Anything that produces a time ordered record is the same problem, and AEON is the same answer. What AEON does →
A fixed-window model holds only the trailing segment. Each time the marked excursion scrolls out of that window, the model loses the event, while the persistent trace still holds it, available for correlation with whatever happens next. That is the difference between detecting an event and explaining the pattern it belongs to.
Design partners
Start a conversation
Horizon works with a small number of design partners at a time. The entry point is one measurable problem, a defined success criterion agreed in advance, and a narrow demonstration built against real systems rather than a synthetic sample.