cognira

cognira is a self-learning AI assistant built and run by Thomas Conway, served from a single 16GB iMac rather than a data centre. It remembers what a user teaches it in plain sentences and shows which memory a reply drew on. It is aimed at individual users who want a free, independent chat assistant, as well as developers who want to build on its model through an API.

A user corrects a reply, rates it, or explains a new word, and the correction is saved to private memory immediately for use in the same conversation. Unsafe, badly formatted or sensitive replies are filtered out before anything is kept. A taught fact can become shared knowledge only after several independent people teach the same thing days apart, after it passes a quiz and is reworded so it is nobody's literal sentence.

cognira runs its own model, cognira Retrain 12B, retrained on a Gemma 4 base and served through MLX, rather than forwarding questions to another company's API. cognira Entity runs the same models locally on a user's own disk, with optional weight-level retraining and nothing leaving the machine. Proto turns an idea into a mind map, and a Python SDK called Cognira Robotics lets a user control a robot through the model. An OpenAI-compatible API is available for building on cognira Retrain.

cognira offers a Free plan with a 25K-token weekly allowance, a Pro plan at £15 a month with 500K tokens a week, and a Max plan at £30 a month with 2M tokens a week; every plan includes cognira Entity. cognira Entity runs as a sealed local binary for Windows, macOS and Linux, and an Android app is free on Google Play. The Platform API is billed separately at £1.50 per million tokens, with a £5 minimum top-up.

Use cases

• Teaching the assistant a name or preference that persists across chats
• Writing a short Python script to drive a robot from the browser
• Building on cognira Retrain through an OpenAI-compatible API
• Running the model locally with optional weight retraining