You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$40,000 left to reach the funding goal.
AI assistants and companions are gaining long-term memory. A system that remembers your history and is always available can become very hard to leave, and today nobody tests for that. Kislon.ai builds Liriel, an open cognitive architecture for persistent AI where memory, relationship and motivation are explicit, inspectable mechanisms. This project builds an "exit layer" into it: dependency signals, a protected step-back protocol, and portability of the person's data. It also builds an open test suite that any persistent-AI system can run.
Threat model of emotional dependency in persistent AI and draft decision rules (months 1-2).
Exit layer in the open-source Liriel reference implementation (months 3-5).
Open test suite and a small volunteer study with the Telegram prototype (months 6-7).
Presence Advisory Issue 03 with the decision rules and results (months 8-9).
The test suite runs against the Liriel prototype and reports, for each check, whether stepping away stays possible: the person can export or delete their data, the system accepts refusal without pressure to continue, and leaving does not lose anything the person needs. Success means these checks pass, are published openly, and can be run by other teams on their own systems.
Minimum USD 10,000: exit layer and test suite, part-time. Goal USD 40,000: the full 9-month plan, including founder research time, a part-time contract engineer, API/compute, and the volunteer study.
Michelle Rodrigues Martins, psychologist and self-taught AI researcher, founder of Kislon.ai. Author of the Liriel specification (Volume 1, three public revisions), the MOV and Ordinance open data matrices and Presence Advisory Issues 01-02, all at https://kislon.ai. Code: https://github.com/kislonliriel.
The most likely failure is limited engineering time: the founder works alone and part-time. Even then, the threat model and decision rules would be published openly and remain useful to other teams.
None. An application to the Foresight Institute (AI for Science & Safety Nodes) was submitted in October 2026.