Experimental · local-first · owner reviewed
Your chat history is not a brain. It is evidence.
Turn selected WhatsApp conversations into inspectable style, memory, and relationship-aware state without pretending the model is the person.
01Selected sources
02Visible uncertainty
03Owner judgment
Two honest scopes
More data is not automatically more truth. Start with the representation you can actually test.
A / narrow
How does this person tend to write here?
One selected participant, usually one conversation, converted into relationship-scoped style examples and held-out evaluation.
B / multi-chat
Which patterns survive across relationships and time?
Selected owner messages plus an optional interview become versioned evidence, claims, relationship profiles, uncertainty, and typed state.
Evidence before imitation
The canonical state sits outside the model. It can be inspected, dated, scoped, corrected, and evaluated.
Choose local chats explicitly. Nothing silently expands to all conversations.
Owner evidence stays distinct from third-party context and cannot borrow its identity.
Patterns carry time, relationship, confidence, provenance, and sensitivity.
Coverage gaps become owner questions instead of confident invention.
A correction creates a successor and leaves the historical record visible.
Eight axes stay separate, with privacy failures blocking an overall pass.
The guided run
They are stages and receipts, not twelve models and not a magic replica score.
Read the system
Start with a runnable path, then inspect exactly how extraction, typed state, correction, and evaluation work.
Open documentationExperimental means bounded
Not a complete person Chats are partial, selected, relationship-specific evidence.
Not autonomous The system cannot send, promise, transact, or represent the owner.
Not finished Automatic typed-state extraction and complete deletion propagation remain incomplete.
Not an official API WhatsApp for Mac is version-sensitive; wacli is a third-party mirror.
Private by default