# Mura ALife Labs — Research (Project NeuroSeed) > Project NeuroSeed is the research program of Mura ALife Labs. It asks whether intelligent systems can learn as well as today's neural networks while costing a tiny fraction of the compute, learning continually, and forgetting exactly on request. - **Organization**: Mura ALife Labs (Independent Research Lab, founded 2026) - **Principal Scientist & Founder**: Prince Siddhpara - **Website**: https://research.mura-alife.com/ - **Main Portal**: https://mura-alife.com/ - **About**: https://about.mura-alife.com/ - **Contact**: contact@mura-alife.com ## Core Research Thesis Modern deep learning relies on brute-force scale, trillion-parameter matrix multiplications, and massive GPU data centers. Mura ALife Labs develops alternative foundations for artificial life and cognition: 1. **Low Compute**: Systems designed to train and infer on commodity CPUs and micro-watt neuromorphic chips. 2. **Continual Learning**: Instant learning from streaming data without catastrophic forgetting or full retraining. 3. **Exact Forgetting**: Deterministic, token-level unlearning in microseconds per token, satisfying strict privacy and unlearning requirements. 4. **Biological Realism**: Learning rules compatible with biological synapses, sparse spike timing, and synaptic tagging. ## The Five Research Tracks ### 1. Ikigai (Track 01 — Active) - **Domain**: Unconventional Language Intelligence - **Core Concept**: Language engine that learns directly from n-gram counts stored on disk. - **Key Results**: - Memory footprint: Under 1 GB RAM while running. - Unlearning: Exact deletion of any trained article in microseconds per token. - Performance: Outperforms transformer baselines on 2M-token (+23.6%) and 12M-token WikiText-103 slices. - Hardware: Operates natively on standard consumer CPUs without GPUs. ### 2. Musubi (Track 02 — Active) - **Domain**: Neuromorphic Hardware Computing - **Core Concept**: Pairs host-side count structures with a sparse spiking neural state running local synaptic learning rules. - **Key Results**: - Operates end-to-end within the strict memory and precision budget of a single neuromorphic chip. - Evaluated on WikiText-2, surpassing state-of-the-art published spiking language models. - Exact unlearning of training documents on command. ### 3. Musubi Bio (Track 03 — In Progress / New) - **Domain**: Computational Neuroscience & Simulated Living Neurons - **Core Concept**: Translating Musubi's learning principles into biological neuron simulations (Brian2, NEST) with synaptic tagging and capture (STC). - **Key Results**: - Requires ~65× fewer synaptic updates and ~50× fewer operations than digital chip equivalents. - Pure inhibitory homeostatic circuit maintains a stable sparse code (~64 active cells out of 2,048) with zero matrix multiplication. - Architecture prepared for live wetware / living-neuron biological hardware interfaces. ### 4. Yomi (Track 04 — Paused, resumes Feb 2027) - **Domain**: Visual Memory & Character Recognition - **Core Concept**: An editable visual glyph memory on disk instead of a GPU-trained neural network. - **Key Results**: - Learns new fonts and scripts in seconds on a CPU from just a few lines. - Outperforms Tesseract 5 on unseen fonts (2.98% vs 4.78% error without labels; 1.38% with 10 labeled lines). - Outperforms CRNN and Transformer-CTC on IAM handwriting using 1/64th of training data. - Exact document removal and full interpretability back to stored glyph prototypes. ### 5. Soma (Track 05 — Paused) - **Domain**: Embodied Cognition & Physics Simulation - **Core Concept**: Autonomous embodied agent exploring 3D MuJoCo physics environments without artificial neural networks. - **Architecture**: Coupled rotating oscillators using algebraic bind/unbind operations to unify locomotion, factual memory, and motor skills. - **Key Results**: - Zero-shot spatial shortcut navigation (0.26° heading error) without trajectory search. - Robust to sudden morphological change (limb elongation or loss mid-stride) without controller retraining (91-129% gait retention). - One-shot motor skill composition (0.000 error). - Open-ended evolutionary morphology: 4,685 creatures across 4,623 distinct body plans in 385 generations. ## Full Technical Documentation See [llms-full.txt](https://research.mura-alife.com/llms-full.txt) for detailed mathematical and experimental documentation.