MnemonicAi is a custom model with a brain-modeled memory system. Hippocampus-style working memory gates what matters, and reinforced memories get baked into the weights themselves. It runs on your hardware, under your control.
Point it at your work. Come back tomorrow — it still knows.
# any OpenAI-compatible client works
client = OpenAI(base_url="https://api.mnemonicai.org/v1")
Every conversation starts from zero. Stuff the history back in each time, pay for it each time, and watch it fall off the end anyway.
Retrieval bolted onto a model that never learns. The memories sit outside the mind — recall is a database query, not knowledge.
A mind doesn't reread its diary every morning. It consolidates.
Every message runs through a hippocampus-style pipeline — perceive, gate, hold in working memory. Noise never makes it past the gate.
Reinforced memories are periodically baked into a QLoRA adapter on the base model — recall persists even with the memory system detached.
Replay buffer, eval-based stopping, adapter versioning with rollback. It learns from you without drifting or overfitting to noise.
Memories that never get recalled fade — the way they should. Forgetting is a feature, not a failure mode.
Self-hosted on Ubuntu 24.04 with Docker or K8s. Two GPUs split the work — one for inference, one for training. Nothing leaves your machine.
Advanced reasoning, vision, tool use, and strong physics and math — on the fine-tuned ornith-1.0-9b base. Memory doesn't cost you capability.
Point any OpenAI-compatible client at the hosted API, or run the whole thing yourself:
# hosted base_url = "https://api.mnemonicai.org/v1" # self-hosted base_url = "http://127.0.0.1:8400/v1"