← Versida / Inside
The main page is for the feeling. This one is for the mechanics: how a reply is put together, what a memory carries, what we've actually verified, and where we still need you to tell us we're wrong.
01 The short version
A language model, wrapped in a character and memory system. The interesting engineering is the application layer around generation: authoring, context selection, persistence, retrieval, controls and verification.
02 One reply, end to end
Each turn the app reads the situation, retrieves relevant evidence, selects character, world, memory and open-plan context, then runs visibility, provenance and budget checks before anything reaches the model.
Retrieval uses semantic representations and graph relationships to offer evidence. The response model still decides how to express it — retrieval success alone doesn't guarantee a correct reply. After streaming, eligible evidence is processed for later. That background work is scheduled application work, not a companion thinking while you're away.
The system prompt is composed from separate rooms. Stable identity first, fresh moment last. Only the parts that fit the moment are selected, rather than placing the full biography into every reply.
08 What we won't claim
Also off the table without separate evidence: end-to-end encryption, complete deletion, exclusive model access, no third-party processing, a monthly cost, instant latency, clinical benefit, or being better than anyone else.
members only
Join to read the schema, the memory model, the evidence ledger and the open questions we want your critique on.