Active product · Sign-in required
Lexi
The conversation shouldn’t start over.
A companion is a demanding test of continuity. A conversation can be convincing in the moment, but the experience changes if the next session forgets what mattered—or if changing the model changes who you seem to be talking to.
I built Lexi around that continuity problem. She brings together a persistent identity, personal memory, and a visual presence, with the aim of making conversations feel connected over time.
That pushed the work beyond dialogue generation. Memories need to belong to the right account. Corrections need to carry forward. Background reflection needs to stay connected to what was actually said. Avatar generation needs to fit into the experience without holding up the conversation. Identity and personality need a home outside any individual model response.
Lexi is the most productized of these projects: an application people can sign into and interact with. She also gives the broader work a human measure of success. Persistence matters because of what it lets someone experience when they return.
Continuity has an owner
Canonical personal memory belongs to the signed-in account and lives in PostgreSQL. Background reflection draws on conversation records; the reviewed deployment retained memory and reflection delivery through a backend replacement and checked isolation between accounts.
Persona controls live outside an individual model response. Avatar generation runs asynchronously through a separate image pipeline, so creating a visual identity does not have to block the conversation.
Muninn contributed to the memory architecture historically. The reviewed implementation disables its shared personal-memory projection; it is not Lexi’s current canonical memory store. Reflection quality remains under development, and device push delivery is not verified.