About AwareLiquid

What This Is

AwareLiquid is an open research framework, not a commercial product. It organises models into two series: M-Series (M1 — cognitive, slow-thinking, GWT + sleep consolidation + EWC) and O-Series (O1 — edge, fast-thinking, continuous-time liquid ODE with constant-memory state). Both series build on the MT-LNN backbone — a continuous-time recurrent network whose hidden state evolves according to learnable time constants — extended with neuroscience-inspired modules.

Research Focus

AwareLiquid does not claim to match frontier models on general dialogue, reasoning, or benchmark scores. The live demo has two distinct models: M1 (the large, cognitive model — a liquid-core adapter, a fraction of a percent of the weights) on a frozen TinyLlama-1.1B-Chat base for bilingual Q&A — general language ability comes from the base, GWT + continual-learning mechanisms from the adapter — and O1 (the small, edge model), a 48M from-scratch model with no base, the native architecture for hardware / always-on inference. Both are research artifacts, not general assistants. The project competes on a different axis:

• Frontier LLMs: weights frozen after training; adapt via context or retrieval; the underlying model no longer learns.
• AwareLiquid research: explore weight-level continual learning and anti-forgetting mechanisms — elastic weight consolidation (EWC), Hebbian plasticity, memory reconsolidation.

Current Status (Honest)

• Built MT-LNN architecture, all liquid core modules, native inference server — this site's demo runs on it.
• Built Continual learning experiment framework + EWC consolidation mechanism, fully reproducible.
• Built Hebbian plasticity ablation suite — initial results show the term is currently inert at tested scales.
• Built Spatial computing modules (causal steering, associative memory, grid-cell scaffolding).
• Built Automatic episodic conversation memory ("remembers you") with typed cards — identity / preference / emotion / plan / relationship / health / detail — over a multilingual sentence encoder. Honest scope: this is RAG-style recall of what you actually told it, plus a lightweight semantic tag; it adds no understanding, reasoning, or emotion.
• Target Achieving "usable" old-domain retention after continual learning (currently ~7.6× PPL degradation; threshold is ≤5×).
• Target Larger-scale model with better generation quality and validated spatial advantages.

For full experimental results, methodology, and honest verdicts on each claim, see the Research page.

Intellectual Lineage

MT-LNN builds on liquid time-constant networks (Hasani et al., MIT CSAIL / IST Austria), which showed that continuous-time RNNs with learnable time constants are expressive, stable, and interpretable. AwareLiquid extends this foundation with modules drawn from neuroscience — global workspace theory (Baars/Dehaene), predictive coding (Friston), and Hebbian plasticity — to explore whether such architectures can support lifelong learning.

Open Science Commitment

All code, experiment scripts, and reproducible configurations are on GitHub. Every result on the Research page includes methodology, seed counts, statistical tests, and an honest verdict. Results that do not yet meet the threshold for "supported" are labeled accordingly — we do not round up preliminary trends into success claims.

Author & Contact

AwareLiquid is an independent research project. The codebase, experiments, and this website were designed and built by a single researcher as an open-science contribution to the continual learning community.

For collaboration inquiries, reproducibility questions, or research discussions, open a GitHub issue or discussion thread. All communication is public and on-record for open science.