Return
01Pick up the thread.
Relevant memories, unfinished topics, tone, and prior boundaries can shape what happens next instead of disappearing at the end of a session.
Xiara
A conversational memory companion built to continue, not reset.
Xiara is a working conversational product for people who want a conversation that can keep its thread over time. You can talk normally, decide what should be remembered, inspect or change memory, and return later without pretending the previous conversation never happened.
The model writes the reply. Xiara carries what should survive between conversations.
Current product
- Text
- Working today
- Memory
- Explicit + user-controlled
- Voice
- Next active frontier
Using Xiara
Come back without starting over.
The basic interaction is simple: talk, leave, return. What makes Xiara different is what happens between those moments — what carries forward, what stays private, and how the next conversation changes because the previous one happened.
Working productXiaraXiara is live as a text-first product. Open it directly here.
Open xiara.coMemory
02See what carries forward.
Memory is explicit. You can ask Xiara to remember something, inspect it, edit it, pause it, or delete it instead of relying on an invisible profile.
Rhythm
03A conversation should not feel like a questionnaire.
Xiara should be able to observe, suggest, follow a thread, leave space, and sometimes stay quiet rather than answering every moment with another question.
Private expression
04Familiarity changes how people speak.
Shorthand, ambiguity, contradiction, adult language, and unfinished thoughts become part of the product problem once the conversation has history.
Under the hood
The model writes the reply. Xiara carries what came before.
The conversation model is replaceable. Memory, context assembly, delivery, boundaries, and surface behavior live outside it so a provider or model change can improve capability without turning Xiara into a different product overnight.
Current product work
The next problems are product problems.
The roadmap is not a pile of features. It is a set of behaviors that have to work together: useful memory under user control, a conversation that still feels like itself after a model change, voice that understands turn-taking, and initiative that adds movement without becoming pressure.
See the research behind these choices- 01
Memory should become more useful without becoming invisible capture.
- 02
Voice has to understand pauses, interruption, and who owns the turn.
- 03
A model change should improve capability without making the conversation feel reset.
- 04
Initiative should add movement without becoming pressure or attention-seeking.
Design boundaries
What Xiara is not trying to optimize.
A long-term conversational product can become manipulative quickly if the wrong metrics win. A shared history matters, but not at the cost of autonomy or user control.
- 01
It is not a therapist or a clinical substitute.
- 02
It is not an engagement machine optimized to keep you talking longer.
- 03
It is not trying to remember everything you say.
- 04
It is not designed to maximize disclosure or attachment.
- 05
It is not tied to one model, provider, interface, or future device.
Surfaces
Text works now. Voice changes the product.
Text is the current working surface. Voice is next because spoken conversation introduces timing, silence, interruption, turn ownership, and delivery choices that do not exist in the same way on a screen.
Now
01Text
The current product supports conversation, explicit memory, and precise user control in a surface that remains easy to inspect.
- Conversation
- Memory controls
- Inspectability
Next
02Voice
The next active frontier adds pauses, interruption, turn-taking, spoken delivery, and the problem of deciding when the system should speak at all.
- Turn-taking
- Interruption
- Spoken delivery
Exploration
03Ambient presence
Possible hardware explores a more persistent physical surface, but it remains optional. Xiara has to work as a relational system without depending on a device in the room.
- Listening boundary
- Visible state
- Optional hardware
Voice + ambient presence study
What still has to work
A long-term product has different failure modes.
01
Identity through model changes
A provider upgrade should improve capability without making the conversation feel like a different mind took over.
02
Useful memory without overreach
Xiara has to become more specific over time while keeping memory visible, selective, and under user control.
03
Initiative with restraint
Knowing when to move the conversation matters as much as knowing when silence is the better behavior.
04
Voice without surveillance
A spoken or ambient surface should perceive only what it needs and make it clear when it is listening.
Xiara is the working product: a conversation that can remember, adapt, and continue without hiding how that thread is maintained.