Hi HN — I built this. ConwAI was the alpha name. Cognira is the live product. It’s a small self-learning system that can pick up a correction, a rating, or a short explanation in the same conversation. Confirmed meaning goes into private memory immediately. Sensitive, badly formatted, or down-rated replies never enter the learning queue. Weight updates only run when there’s a clean batch, the server is idle, and a few checks pass. One bad chat cannot rewrite the live model. The hosted chat is not a wrapper. I train and serve the models myself: cognira Retrain 12B and Omega 188M, MLX inference, on a living-room iMac. There’s also cognira Entity — local inference, memory, and optional weight retraining on your disk. Nothing leaves the machine. Free hosted chat, no card. OpenAI-compatible API at api.cognira.dev (£1.50 / 1M tokens, £5 minimum top-up). What I’d like feedback on:

whether the “teach once, use immediately” memory actually feels useful whether the guardrails are too tight or not tight enough whether a 12B + 188M pair on one Mac is an honest local story or just underpowered

Site: https://cognira.dev

Local app: https://cognira.dev/entity

API: https://cognira.dev/platform Happy to answer questions about the learning loop, the queue, or the MLX setup.