Onyon SystemsISOLATION
Separate personal contexts
A person’s records belong in their own context, with boundaries between accounts.
Private memory.
Persistent context.
We are building the memory foundation for privacy-first AI agents. Zelo is the first application; continuity across other apps and channels is our longer-term vision.
01 / Privacy
Built to be checked.
Protection is a system of decisions: who gets access, how long data stays, and what a model is allowed to learn.
01 / PERSONAL CONTEXT
Start with the person
The agent’s memory should serve the person it belongs to, with clear boundaries around access and use.
02 / DATA
Design for control
What is kept, what reaches an AI provider and what can be removed are decisions at the heart of the product.
03 / DEVELOPMENT
Be clear about the stage
Zelo is in development. Bloom and connections to external assistants and channels are planned.
04 / Limits
Establish the boundary
Participation in an ecosystem should come with explicit boundaries for what each participant can access.
05 / CONTEXT
Understand what is shared
A useful response needs context. Explain which information is used and why it is relevant.
06 / CONTINUITY
Keep the person at the center
Our vision is for private memory to remain useful as a person moves between apps, assistants, messages and calls.
02 / The Ring — Persistent memory
A memory with
structure, in layers.
A useful AI product can carry relevant context from one interaction to the next. Facts remain connected to sources, so people can understand and correct what is remembered. Facts, sources, dates, documents, and what is due next. The Ring keeps a structured memory of the subject your companion serves.
Every fact is a line you can read, dismiss, restore, and export.
The memory foundation is under development for Zelo. The broader architecture describes where we want to take personal agents.
FIRST INTERACTION
Establish context
Start with the subject, the question and the information a person provides.
OVER TIME
Keep useful facts
Carry relevant records and their sources forward instead of relying on a single conversation.
NEXT INTERACTION
Retrieve what matters
Use that history to inform the response, while keeping the record open to inspection.
SESSION
- HOURS
- context for the moment, cleared when the moment passes.
RECENT PATTERNS
- WEEKS – MONTHS
- routines, changes, the shape of now.
PERMANENT ARCHIVE
- YEARS
- The record of a life: Inspectable, exportable, and only ever yours.
03 / How it works
Our design principles
The principles guiding our development: private personal context, useful memory and clear boundaries. Availability will be described for each product as it launches.
DATA LIFECYCLE
Memory under your control
Design memory so the person can understand what is kept and how to correct or remove it.
MODEL TRAINING
Personal context stays personal
Our commitment is to build personal memory for the person, without using it to train shared AI models.
USE POLICY
Care is the purpose
Participation by a business should have a clear purpose and permission, with access boundaries around personal records.
OBSERVABILITY
Operational boundaries
Supporting a product should not mean unrestricted access to someone’s personal history.
DATA PROTECTION
Protection throughout
Consider how information is protected when stored, transferred and used to produce a response.
ACCURACY
You confirm what we read
Information read from a document should be reviewable before it becomes part of the record.
TRANSPARENCY
How AI fits in
Explain the role of external AI providers and the context a response may require.
PORTABILITY
Beyond one interface
Connections to assistants, messaging and voice are planned paths for the same private memory.
04 / Technology, privacy and security
Clear about what comes next.
- Is conversation data used to train shared models?
- No. Conversation data is never used to train shared AI models.
- Does context reach a model provider?
- Yes. Selected messages, profile information, relevant facts and document passages may be sent to a provider. The current system does not conceal all prompt content from that provider.
- How does privacy guide the design?
- We start with personal context, clear access boundaries and control over what is remembered. Product-specific details will be published as development progresses.
- Will the agent work outside an Onyon app?
- That is our vision. Connections to AI assistants, messaging, SMS and voice are planned, with no availability date announced.
- What is available today?
- Zelo is in development. Bloom, partner record connections and external channel integrations are planned. We are discussing partnerships ahead of launch.
- What does persistent memory contain?
- Persistent memory brings together useful facts, records and their sources over time. Zelo applies that idea to the history of a pet.
- How can a partner participate?
- We are seeking partners to offer Zelo to their customers, connect clinics and pet shops to Bloom, or explore both paths in the same pilot.