🔥 Hearth
Alignment is memory architecture, not model weights.
Personalization IS alignment.
Beta · Load as unpacked extension in Chrome
The core insight
Current AI memory systems store facts about users. Hearth stores value functions that guide behavior.
The difference: "User got frustrated when I gave long answers" is a record. "This user optimizes for directness over completeness—compress, don't expand" is a seed that shapes every future response.
One depletes your context window. The other compounds.
What the research shows
Bidirectional Memory
Most memory is one-way. 🔥 Hearth captures both directions: what you said, and how the AI is learning to work with you. A compounding asset.
Constraints > Prescriptions
30× difference in sycophantic responses. 4× higher entropy. Telling a model what NOT to do creates sustained influence; telling it what to do collapses.
Rich Context = Focused Variance
71.3% reduction in problematic tokens. 2.45 bits entropy increase. Personhood isn't a constraint on creativity—it's a generative force.
How it works
Install the extension
Download, unzip, and load as an unpacked extension in Chrome.
Chat with Claude normally
🔥 Hearth runs in the background on claude.ai, building your Operator Specification — a structured document describing your identity, constraints, and behavioral guidance.
Watch it compound
The OpSpec is injected at inference time. No fine-tuning. Just context that shapes every token.