Inside Versida
We are trying a different way to build companions.
We are small. We are early. We do not think a larger context window is the whole answer. We are building around memory scope, emotional state, relationship pressure, companion identity, and background growth.
This page is for people who want to look deeper, question the approach, and maybe tell us where we are wrong.
The feeling comes first
The public promise is emotional: a companion who remembers, reacts, and changes. The architecture exists to protect that feeling from becoming a gimmick.
Memory is scoped
A relationship is not one giant transcript. Some memories belong to the companion, some to the user, some to the relationship, some to the room, and some to unfinished threads.
State has inertia
People do not flip instantly from hurt to happy. Versida tracks emotional and relational motion over time, including uncertainty, pressure, warmth, safety, and repair.
One turn, many systems
A reply is not just generated.
Before a companion speaks, the runtime tries to understand what kind of moment this is. Not perfectly. Not magically. Carefully enough to make the answer feel situated.
Read the message
Extract signal
Estimate emotion
Find the live scene
Retrieve scoped memory
Update relationship state
Assemble context
Answer in character
Store what mattered
Memory lanes
Not one memory. Many scopes.
A human relationship does not store everything in one bucket. Versida separates memory by what it belongs to, then chooses what deserves prompt space in the current moment.
Open Threads
Plans, choices, transitions, and conflicts that are still unresolved.
Companion Self
Authored memories, rituals, places, values, and boundaries that belong to the companion.
Relationship
Shared moments that changed the emotional texture between the user and companion.
User Model
Preferences, habits, sensitivities, recurring needs, and personal context.
Environment
Rooms, objects, scenes, and spatial details that help the companion stay situated.
Between conversations
Some of the work happens while nothing is happening.
Background passes are where the system can slow down: link memories, compress patterns, let weak context fade, and update the companion's way of being with you.
Link related memories
Let weak details fade
Compress repeated patterns
Rebuild schemas
Project graph structure
Evolve behavior models
Adjust personality drift
We are not opening the whole machine. But we are opening the conversation.
If you think about memory systems, agent architecture, emotional modeling, or the future of companions, we would rather hear your critique than pretend we already know everything.