A self-learning AI assistant built and run by a single person (Thomas Conway) on one 16GB iMac, not a venture-backed lab. It runs on a 12B model (cognira Retrain 12B) based on Gemma 4, served via MLX. Its core differentiator is a three-layer learning system: instant private memory (teach it a fact or preference and it applies to your next reply, with a one-tap Forget), guarded shared knowledge (only promoted after multiple independent users teach the same thing, pass safety screening, and survive a quiz—every rejection is public), and weight-level retraining during idle time. A local version, cognira Entity, is a sealed binary for Windows/macOS/Linux that runs the same model on your own hardware with optional local retraining. Pricing is simple: Free £0 (25K tokens/week), Pro £15/month (500K tokens/week), Max £30/month (2M tokens/week); every tier gets the same model and features, only the weekly allowance differs. An OpenAI-compatible API is available at £1.50/1M tokens .
Pros & Cons:
Pros: The self-learning loop is unusually transparent—you can see exactly which memory shaped each reply and forget it instantly, and the public rejection ledger at cognira.dev/learning shows what was refused and why, which is rare honesty ; the local cognira Entity binary means the same models can run entirely on your own machine with no data leaving, included on every plan including Free ; pricing is genuinely simple—Free/Pro/Max differ only in weekly token allowance, not in model quality or features ; an OpenAI-compatible API at £1.50/1M tokens gives developers a cheap way to build on it ; the founder is solo and self-funded, which is why the product optimizes for low-energy serving on one machine rather than GPU clusters .
Cons: Solo founder operation with one 16GB iMac serving the hosted product raises serious questions about reliability, uptime, and scaling—this is a workshop, not a data centre ; the product is very new (launched September 2026) with no independent third-party reviews (G2/Capterra/Trustpilot absent), only a TAIFY listing that estimates ~$20/month typical use ; analytics and session replay are extensive—PostHog records complete prompts and replies, linked to your account, with no in-app opt-out (the terms say the self-learning system needs this data) ; the hosted model runs on consumer hardware (one iMac), so response speed and concurrency are inherently limited; no SOC 2, HIPAA, or enterprise compliance surfaced—it is explicitly a solo indie project, not enterprise-ready.
Who is it for?:Privacy-conscious users who want an AI that learns their preferences and can run fully locally via cognira Entity; developers curious about a transparent, inspectable self-learning loop with a public rejection ledger; users comfortable with a solo-founder project who value honesty over enterprise polish; budget-conscious users who want a capable local model without paying per-token cloud fees.