cognira is one person, one iMac, and a model that learns in the open.

The person who builds them is still Thomas Conway. The names you will see on this site:

• Omega — the original 188M-parameter checkpoint. Deprecated in-product, still open-source on Hugging Face.
• cognira Retrain 12B — the current model, retrained on top of a Gemma 4 base and served through MLX on our own machine.
• cognira Entity — the fully local version that lives on your own hardware and learns you, with nothing leaving the machine.
• Platform API — an OpenAI-compatible endpoint so you can build on the model from your own code.

Which Conway?

Not the mathematician, not the Game of Life, not the singer, and not a town. If you searched “Conway” looking for an AI that learns from you and shows its working, this is the one.

cognira is a solo, independent build. The model retraining, the MLX serving stack, the memory system, the verification gate, the public ledger, the website and the emails from [email protected] are all one person. There is no marketing team, which is why the copy on this site says “we do not know” where a bigger company would have invented a number.

That constraint shapes the product. A single 16GB iMac is not a data centre, so efficiency is not a slogan here — it is the only way the thing runs at all. That is where the eco AI principles come from. And with one pair of hands, the only sustainable way to make the model better is to let the people using it teach it — carefully, visibly, with a gate. That is where the self-learning loop comes from.

It learns from the people who use it

Teach it a preference, a fact or a word and it is written to your private memory immediately — the next reply that uses it shows you exactly which memory, with a one-tap Forget. A taught word can go further, but only after several independent people teach the same thing days apart, it is reworded so it is nobody’s literal sentence, and it survives a quiz. Every rejection is public at cognira.dev/learning.

Our serving stack. Our hardware. A replaceable checkpoint.

Inference is served by cognira, not forwarded to another model API. The checkpoint sits behind an interface so memory, safety rules and user ownership survive future model changes. This is a workshop, not a monument.

Efficiency we can measure, not invent

One 16GB machine instead of a GPU cluster. Response budgets that stop needless token generation. Background learning that defers when inference is busy. And no carbon number until we can meter it properly.

The compact local serving setup keeps costs low enough that the Free plan can include the full model and instant private memory with no credit card. Pro and Max raise your weekly allowance and help pay for serving and the guarded learning loop. Every tier gets the same answers.

Conversations are stored so eligible parts can enter the learning process. There are no ads and your data is not sold. If you want your account and chats removed, you can delete them anytime — see the privacy policy.