Constant memory.
Unbounded context.
AwareLiquid is a bio-inspired cognitive AI framework. The attention-free O-Series carries a fixed-size state instead of a growing KV-cache, so context length stops dictating memory.
O-series state vs. KV-cache
flat from 512 to 1M tokens
fast-weight vs. attention/LoRA
attention-free O-series
Three Mechanisms That Make Thinking Possible
AwareLiquid stacks three neuroscience-derived subsystems on a continuous-time liquid core — each solving a distinct cognitive problem that Transformers and SSMs cannot address.
Global Workspace Theory (GWT)
Specialist cortical modules compete for a shared broadcast channel — the "global workspace." The winner propagates its representation to all other modules, enabling selective attention, emergent integration, and conscious-like information routing.
Sleep Consolidation (NREM Bridge)
Synthetic NREM cycles replay short-term session activations and compress them into a persistent graph-memory structure. Synaptic Homeostasis (SHY) downscaling prunes redundant connections, boosting signal-to-noise. The model grows smarter during rest.
Liquid ODE Dynamics
Hidden state evolves as a continuous-time differential equation with learnable time constants — inspired by microtubule dynamics. No discrete layer stacking, no KV-cache growth. Constant-memory state enables edge deployment at millisecond latency.
Built for streams that never stop, on hardware that cannot grow.
Three deployments with something real behind them — measured, shipped, or both. Every number links to open code; anything without data says so.
Battery management
State-of-health regression on voltage/current/temperature streams, on the pack controller itself.
Measured — NASA PCoE cells +7.7% at 80% sample loss vs +31% LSTM/GRU · 2.6 KB streaming state. 10 seeds, whole cell held out.Hearables
Always-on wake-word spotting where the microphone never closes — one keyword, measured in English. Open source as O1-Sound.
Measured — single keyword only English “hello”: FRR 0.14 ± 0.05 at FAR 0.05 across 3 seeds. 5.03 MB fp32 / 1.27 MB int8, 5,120 B constant state. Not production grade: seed-to-seed spread is wider than any change we have tried, and the multilingual version measured worse than never firing — one binary class cannot span greetings that share no phonetics.Financial documents
QA over filings far longer than the context window, with the base model frozen behind an API.
Measured — M2 adapter 44/48 (91.7%) across five financial-document domains · ~2.8k tokens per question. A regression baseline, not a leaderboard.Also targeting In-vehicle systems · On-device assistants · Industrial robotics — architectural fit (bounded state over an unbounded session, native Δt), no pilot data yet.
Measured.
Reproducible.
Every claim below is measured on real runs and backed by open code and data.
| Capability | Result |
|---|---|
| O(1) inference memory (O-series) | 0.381 MB flat → 8063× smaller @1M tokens |
| Cross-window associative recall | 0.56 vs 0.000 (attention/LoRA) |
| Cross-session snapshot → restore | bit-exact lossless |
| Stable training at 125M scale | no NaN, 20K steps, 3 seeds |
| Irregular-sampling robustness (battery SoH) | 4× steadier than LSTM/GRU when samples drop |
Full methodology, raw data and reproduction scripts: BENCHMARKS.md.
Try Both Models in Your Browser
Two distinct models, both live. M1 (large, cognitive) runs trained liquid-core residual adapters on a frozen TinyLlama-1.1B-Chat base — a general assistant. O1 (small, edge) is the native 48M liquid architecture trained from scratch with no base model — the raw backbone for hardware / always-on inference. Switch between them below. Ask in English or Chinese.
Research-grade demo — quality is experimental. What to expect →
A Native Substrate for Cognitive Agents
Transformers excel at static retrieval. AwareLiquid is built for something harder: agents that form episodic memories across sessions, coordinate through shared global workspace broadcasts, and continuously consolidate knowledge without catastrophic forgetting. The foundation for an Awareness Market of collaborating cognitive agents.
See current architecture →Same Lineage. Different Frontier.
Both projects build on MIT liquid neural networks. Liquid AI solved static-weight inference efficiency. AwareLiquid asks the next question: can a model acquire new knowledge without erasing the old?
Efficiency
- SSM + linear attention + MoE hybrid
- 350M – 24B parameters, commercial scale
- Weights frozen post-training; in-context only
- Targets industrial time-series & edge devices
& Lifelong Learning
- Continuous-time RNN + GWT + Sleep + EWC
- 48M – 60.7M, open research prototype
- Online weight updates — learns without forgetting
- Targets episodic memory, agent cognition & adaptation