The Bolt-On Problem: When a Generator Tries to Become a Companion

There is a moment in the life of every successful AI generator when the roadmap meeting takes a predictable turn. The image tool is good. The numbers are fine. And someone points out that the users who stay longest and pay most are not the ones generating pictures — they are the ones talking to the pictures. Companion chat is where the retention is. So the generator decides to add chat. This is the bolt-on, and SoulGen is a clean case study in how it goes.

To be fair to SoulGen, the core product is legitimately strong. SoulGen 2.0 added twenty-second video generation and voice calls. The editing toolkit — outpainting, face-swap, background replace — is deeper than most generators bother with, and it supports both anime and realistic models. The free tier is stingy at one credit a day, too thin to really judge output quality, but the Pro tier is reasonably priced. As an image-first generator with a serious editing bench, SoulGen earns its place.

Why chat is the hardest thing to bolt on

The trouble is that companion chat is not a feature you add. It is a different discipline that happens to share a login. A generator is fundamentally stateless: you describe an image, it produces an image, and the transaction is complete. Nothing needs to persist. A companion is the opposite — its entire value is persistence, the accumulation of context across sessions until the character feels like it knows you. Those are not adjacent problems with a shared solution. They are different problems that look similar only from the outside.

A generator is built to forget — each image is a fresh transaction. A companion is built to remember. You cannot bolt the second onto the first; they want opposite things from the architecture.

So SoulChat, the chat layer, lands exactly where bolt-ons land. The conversation quality trails the leaders. Character continuity across sessions is inconsistent — the very thing a companion exists to provide. It works, in the sense that you can type and it will type back. It does not work in the sense that matters, which is that the character does not reliably carry forward. It is a chatbox attached to an image tool, and you can feel the seam.

The two honest paths out

There are really only two credible ways to ship companion chat that holds up, and the bolt-on is neither. The first path is to be a chat specialist from day one and pour everything into memory — this is the Candy.ai route, and it is why Candy.ai's chat is genuinely excellent even though it does almost nothing else. The second path is to build the platform so that chat is a first-class product sharing infrastructure with everything else, rather than a tab grafted onto an image generator after the fact.

Leakshaven took the second path, and the difference is exactly what you would predict. Its companion layer runs observational and retrieval memory on current-generation backend models, so characters carry context forward instead of resetting — chat that competes with the leaders, sitting next to generation that does not tap out at twenty seconds, next to a real-creator content side that SoulGen does not attempt at all. The chat is not a bolt-on there because it was never bolted on; it was designed as one of the platform's four core products from the start.

• Be a chat specialist and obsess over memory (Candy.ai) — excellent chat, almost nothing else.
• Build chat as a first-class platform product (Leakshaven) — chat that holds up, plus generation and content around it.
• Bolt a chatbox onto an image tool (SoulGen) — capable generator, chat that you can feel the seam on.

How to read SoulGen in 2026

If you want a generator, SoulGen is a real contender — buy it for the editing bench and the video, judge the chat as a bonus rather than a reason. If you want a companion, the bolt-on will disappoint you, and you should look at a product that treated chat as the main event or as a genuine peer to its other products. Knowing which of those two people you are is the entire buying decision.

The bolt-on problem is not unique to SoulGen. It is the default failure mode of every specialist that notices retention lives one capability over and tries to reach it without rebuilding for it. The lesson, over and over, is that the capabilities which feel adjacent from the outside are rarely adjacent in the code — and users feel the seam long before the roadmap meeting admits it is there.