Fourwheels2512/live-continual-learning

🧠 Live Continual Learning — by ModelBrew

A new frontier in machine learning: knowledge that is always current, never confidently wrong, and fully governable. Teach a fact in one forward pass — no retraining, no LoRA, no router — and the model can use it immediately without forgetting anything it already knew.

This Space is an interactive demo you can try right now on CPU.

What you can do here

• —🧠 Live Learning — teach the model a new fact and ask about it instantly. Ask about something it doesn't know and watch it say "I don't know" instead of inventing a confident wrong answer.
• —🛡️ Governed Memory — see everything the brain knows, erase a fact (right-to-be-forgotten) and get a salted, byte-verified certificate of erasure, and inspect the tamper-evident hash-chained ledger (try the "simulate tamper" button).
• —⚔️ vs RAG vs Fine-tuning — an honest, side-by-side capability comparison.
• —📊 Benchmarks — validated, model-invariant results across Qwen3-4B, Llama-3.1-8B and Phi-3-medium-14B.

Core highlight features

♾️ Zero forgetting · 🎯 Zero cross-fact hallucination · 🧩 3 model families · 🚫 No LoRA · 🧭 No router · ⚡ Instant one-pass updates · 🙅 Never confidently wrong (says I don't know) · 🧾 Tamper-evident ledger · 🔏 Proof of erasure · 🧠 Governed memory · 📡 Live learning · 🗂️ Remembers everything.

How honest is this demo?

This public Space is a faithful behavior twin of the ModelBrew Live-CL engine. It reproduces the engine's external contract — teach · ask · abstain · erase · audit — using lightweight, standard-library governance logic so it runs instantly and free on CPU. It contains no proprietary model code. The production engine runs the same contract on real frozen base models (Qwen3-4B / Llama-3.1-8B / Phi-3-medium-14B).

On benchmark framing: a strong dense retriever can match or beat us on raw paraphrase generalization on some datasets. Our moat is self-calibration (never confidently wrong), zero-forgetting by construction, zero cross-fact hallucination, certified deletion, and a tamper-evident audit trail — the properties that compliance, legal, clinical and finance teams actually need.

Learn more at [modelbrew.ai](https://modelbrew.ai).

Run locally

pip install -r requirements.txt
python app.py

If the Space fails to build, bump sdk_version above to the latest Gradio 5.x release (and the pin in requirements.txt to match).