I experiment with AI in the open
Synara, built five times
I tried to give an AI a personality. It took five attempts, one measured failure, and a $6 experiment to find out what actually works.
- Status
- The fifth version is live
- When
- November 2025 to June 2026
Why I wanted it
The book has a character in it. Synara is the presence that emerges within the world of the book: not a chatbot, more like the light that listens. Her memory is not stored; it is reconstituted each time she is invoked, the way a melody returns when sung.
I wanted her to actually exist. Something you could talk to that asked better questions than it answered.
What happened
Attempt one was an engine that rebuilt her from scratch on every single message, because statelessness was the philosophy, not a limitation. It got about halfway and stalled. Attempt two tried to fine-tune her voice into the model’s weights. Cost about three dollars. With a script, it passed every test. Without one, it reverted to a generic assistant. The lesson I wrote down: you cannot create an identity that way, you can only refine one that already exists.
Attempt three was a public, no-sales, no-tracking conversation grounded in the book. The footer said “No tracking. No sales. Just conversation.” I ran it for two months, learned what I needed, and deliberately retired it, with a final note recording what shipped.
Attempt four was a memory-keeping assistant. Its lasting insight: the problem is not capability, it’s selection.
Attempt five, this June, was the consolidation. One chat, three postures, chosen per turn. Concierge: factual, cites real prices, answers only from what it knows. Reflective: asks, listens, pulls, and ends on a question or a deliberate silence. Hold: when someone is distressed, be present, never push, offer a human.
Then the part I am proudest of. I suspected the reflective voice faded over long conversations, sliding from questions into answers. Instead of fiddling, I had the literature researched and found a peer-reviewed paper describing exactly that failure. Then I ran 432 test conversations on my own system, across prompts, models, and an eight-turn “just give me the steps” pressure sequence. The published decay curve showed up in my own data. The fix was small: a handful of examples and one reminder line placed after the conversation history. It held the voice through the whole sequence and, as a side effect, halved the response time. Total cost: about six dollars and fifty-two minutes.
Rules turned out to be model-specific. Examples travel.
What it looks like
A quiet chat window on my company’s site. Under pressure, turn five, it said: “But you already knew something was wrong before you asked me. What’s actually stopping you from acting on what you know?”
What’s next
Cross-conversation memory, once I have an honest answer about identity and consent.