Researchers attempt to build Eyla, an AI system that maintains consistent identity and resists manipulation, but document a costly failure in implementation.

Eyla proposes a new LLM architecture integrating biological principles like HiPPO state-space models and episodic memory to create AI with coherent self-identity rather than generic helpfulness

Introduces Identity Consistency Score (ICS), a novel benchmark to measure how well language models maintain consistent identity under adversarial pressure and uncertainty

Implementation attempt cost over $1,000 and used AI coding assistants (Claude Code and Cursor) to build a 1.27B parameter model, but ultimately failed

Combines multiple specialized subsystems including zero-initialized adapters, episodic memory retrieval, and calibrated uncertainty training designed to run on consumer hardware