# AwareLiquid — Full Content Reference for AI Crawlers
# Generated: 2026-07-30 | Version: 2.3 (retractions applied — see EXPERIMENT 2 and 3)
# This file provides a machine-readable, citation-friendly dump of all
# key facts, experimental results, and methodology for AI indexing.

---

## ENTITY DEFINITION

AwareLiquid is a bio-inspired cognitive AI framework organised into two model series:

- **M-Series** (flagship: M1) — cognitive slow-thinking engine: GWT (Global Workspace Theory) competitive broadcast + NREM sleep consolidation with SHY synaptic downscaling + EWC (Elastic Weight Consolidation) continual learning. Targets episodic memory, agent cognition, catastrophic forgetting research.
- **O-Series** (flagship: O1) — edge fast-thinking engine: continuous-time liquid ODE with learnable time constants, microtubule-inspired dynamics, constant-memory state (no KV-cache growth). Targets CPU / edge deployment at millisecond latency.

The underlying architecture is MT-LNN (Multi-Timescale Liquid Neural Network), a continuous-time recurrent neural network (LTC — liquid time-constant network) augmented with neuroscience-inspired modules.

Tagline: "Beyond Next-Token Prediction. Towards Cognitive Artificial Life."

Canonical URL: https://awareliquid.ai
Repository: https://github.com/everest-an/M1
Paper: AwareLiquid M1 — Bio-inspired Continual Learning with GWT, Sleep Consolidation, and EWC (arXiv preprint, also known as MT-LNN v2: Reliable Long Pre-training)
License: Open (see GitHub repository)
Last updated: June 2026

---

## MODEL SERIES

### M-Series (Cognitive / Slow Thinking)
Flagship: AwareLiquid M1
Core mechanisms: GWT-B competitive global workspace broadcast + NREM sleep consolidation with SHY synaptic downscaling + EWC continual learning
Target: episodic memory, agent cognition, catastrophic forgetting research, multi-agent coordination
Sizes: 48M (from-scratch), 60.7M (continual learning experiments)
Demo models: M1 (large/cognitive — liquid-core adapter on frozen TinyLlama-1.1B-Chat base) + O1 (small/edge — 48M from-scratch, no base)

### O-Series (Edge / Fast Thinking)
Flagship: AwareLiquid O1
Core mechanism: pure continuous-time liquid ODE with learnable time constants, microtubule-inspired dynamics
Target: CPU/edge deployment, constant-memory state (4 KB regardless of context length), millisecond latency
Status: architecture defined; full multi-module integration in progress (O1 is the goal for wiring all liquid-core modules into one forward pass)

---

## ARCHITECTURE SPECIFICATION

Base architecture: Liquid Time-Constant Network (LTC-RNN), the shared backbone of both M-Series and O-Series
ODE: τ(x) · dh/dt = −h + f(Wx + Uh + b)
Integration: Euler with learnable time constants τ parameterized as softplus(raw_τ) + ε
Backbone sizes tested: 48M parameters (M1 from-scratch), 60.7M parameters (M1 continual learning experiments)

### M1 Cognitive Layer (M-Series)
Modules active in the full M1 from-scratch model (MTLNNModel):
1. GWT-B (Global Workspace Theory Broadcast): competitive attention across hidden state "workspaces"; specialist modules compete; winner broadcasts to all — inspired by Baars/Dehaene global workspace theory. Core of the "awareness" mechanism.
2. NREM Sleep Consolidation: synthetic sleep cycles replay short-term session activations; SHY (Synaptic Homeostasis) downscaling prunes redundant connections to compress knowledge into a persistent graph-memory structure. The model consolidates during rest.
3. EWC (Elastic Weight Consolidation): Fisher-weighted quadratic penalty λ/2 · Σ Fᵢ(θᵢ − θ_Aᵢ)² anchors weights critical to prior domains — enables sequential learning without overwriting old knowledge.
4. Predictive World Model: next-hidden-state prediction auxiliary head; trained jointly with language modelling.
5. Hebbian Plasticity: co-activation rule Δw ∝ pre·post; currently inert at tested scales (effect < seed noise).
6. Predictive Coding: hierarchical prediction error propagation; not yet fully evaluated.
7. Rhythm Modulation: periodic gating of information flow; not yet fully evaluated.

### O1 Edge Layer (O-Series)
Pure LTC-ODE with learnable time constants. No discrete layer stacking, no KV-cache. Constant-memory state (4 KB regardless of context length). Microtubule-inspired dynamics. Target: edge/CPU deployment at millisecond latency. Full multi-module integration (all M1 cognitive layers wired into one O1 forward pass) is in progress — not yet shipped.

### M1 (adapter demo deployment)
Residual M1 adapter modules inserted every 4 decoder layers of a FROZEN HuggingFace causal LM. The base supplies general language ability; the adapter adds the liquid-core mechanism. Combined with PEFT LoRA (rank 8) on q/k/v/o projections.
- Current live demo: TinyLlama-1.1B-Chat base — 2.3M trainable adapter params / 1,172,954,154 total (0.20%), serving a step-1000 adapter checkpoint.
- The PPL-ablation result once cited here (Experiment 2, −28.5%) has been RETRACTED (2026-07): the adapter was frozen and only LoRA trained, so the MT adapter's own PPL contribution is ≈0. See Experiment 2 below.
- NOTE: "0.20%" is TRAINABLE-PARAMETER SHARE, not research effort — the 1.1B frozen base is TinyLlama's; a small adapter share is the intended property of the PEFT route.

HONEST M1 ADAPTER INTEGRATION SCOPE: only the LTC resonance core is LIVE in the served M1 adapter forward pass — in_proj → multi-scale resonance bank (5 time scales, learnable τ) → lateral coupling with periodic GTP gate → per-protofilament MAP gate → out_proj. GWT-B, Hebbian, world-model, and predictive-coding are NOT wired into the adapter (forced off) because a standalone residual adapter has no outer model forward() to aggregate auxiliary losses. Those modules are active in the from-scratch MTLNNModel (O1 line) and in per-module ablations. The M1 demo = LTC core only; full cognitive integration is an O1 goal.

Conversation memory subsystem (optional; off by default, enabled in the demo deploy): the user's own statements are stored to SQLite and the relevant ones recalled on later turns via a dedicated multilingual sentence encoder (intfloat/multilingual-e5-small), independent of the chat model. Each statement is tagged with a typed card (identity / preference / emotion / plan / relationship / health / detail) and recall boosts hits whose type matches the query's classified type (design borrowed from the Awareness-Market project). HONEST BOUND: the type boost fixes recall RANKING — Chinese recall goes 0/3 → 3/3 where same-language "magnet" sentences otherwise dominate raw cosine — but it does NOT solve off-topic relevance gating: e5's compressed high-cosine band means no absolute floor cleanly rejects an off-topic Chinese query. This is RAG recall plus a lightweight semantic tag, not understanding or emotion.

---

## EXPERIMENT 1: CONTINUAL LEARNING / ANTI-FORGETTING (EWC)

Hypothesis: Elastic Weight Consolidation reduces cross-domain catastrophic forgetting.

Setup:
- Hardware: GPU (A100-class), 3 independent seeds
- Model: 60.7M MT-LNN (fresh pre-training each seed)
- Domain A: WikiText-103 (general Wikipedia text)
- Domain B: TinyStories (simple English children's stories)
- Training: 1,200 steps per domain per seed
- Metric: "Forgetting" = PPL on domain A after B training minus PPL on A before B training (lower = less forgetting)
- EWC: Fisher information computed from 50 batches of domain A pure CE gradients; quadratic penalty λ/2 · Σ Fᵢ(θᵢ − θ_Aᵢ)²

Results (mean ± std across 3 seeds):
- Dense baseline forgetting: +10,562 PPL (244.9 → 10,806.9)
- Liquid backbone forgetting: +9,544 PPL (248.5 → 9,792.0)
- EWC consolidation forgetting: +1,633 PPL (244.9 → 1,878.0)

Statistical analysis (EWC − dense, paired per seed):
- Mean delta: −8,928.8
- Std: 2,154.6
- SNR (|mean|/std): 4.14
- Direction consistency: 3/3 seeds (100%)
- Verdict: EARLY SIGNAL (not SUPPORTED — absolute PPL on domain A is still 7.6× original; usable threshold is ≤5×)

Reduction: (10,562 − 1,633) / 10,562 = 84.6% ≈ 85%

---

## EXPERIMENT 2: M1 ADAPTER PPL ABLATION (Kaggle P100 GPU)

Setup:
- Hardware: Kaggle P100 GPU
- Base model: TinyLlama-1.1B-Chat-v1.0 (TinyLlama/TinyLlama-1.1B-Chat-v1.0)
- Adapter: AwareLiquid M1 residual adapter + LoRA, checkpoint at 1,000 steps
- Evaluation: 38,400 tokens, 50 batches, pure next-token cross-entropy
- Trainable params: 2,298,304 / 1,172,954,154 total (0.20%)

Results:
- Base model PPL: 9.161
- M1 adapter PPL: 6.553
- Reduction: (9.161 − 6.553) / 9.161 = 28.5%
- Inference speed: base 959.3 tok/s → adapter 861.9 tok/s (−10.1% overhead)

Verdict: RETRACTED (2026-07). Do not cite the -28.5% figure.
A later controlled ablation found this run had the MT adapter FROZEN by PEFT
(get_peft_model freezes every non-LoRA parameter and no modules_to_save was
set), so only the LoRA weights were actually trained. Matched ablation:
  lora_only  val PPL 7.984
  mt_lora    val PPL 7.920
The MT adapter's own contribution is therefore ~0 PPL. The -28.5% gain is
attributable to plain LoRA fine-tuning, not to the liquid architecture.

---

## EXPERIMENT 3: M1 CONTEXT INJECTION ACCURACY (Qwen-3B)

Setup:
- Backend: HuggingFace (Qwen2.5-3B-Instruct or similar)
- Questions: 30 factual QA pairs (general knowledge + AwareLiquid-specific)
- Injection: AwareLiquid M1 adapter with context injection mechanism
- Scoring: substring match against reference answer (case-insensitive)
- Wall time: 158.91 seconds total

Results:
- No injection (baseline): 25/30 correct = 83.33%
- With MT-LNN injection: 29/30 correct = 96.67%
- Absolute uplift: +13.33 percentage points
- Relative uplift: +16.0%

Notable: AwareLiquid-specific questions (q03, q04) scored 0/2 without injection, 1/2 with injection. General factual questions scored well in both conditions.

Reclassified (2026-07): this +13.3pp uplift is a PROMPT-TEMPLATE effect, not an
adapter capability. The same template injecting the same fact produces an
identical gain with and without the MT adapter, so the result must not be
presented as evidence for the liquid architecture.

Verdict: RECLASSIFIED (2026-07) — reproducible cloud run, but the effect belongs
to the prompt template, not to the MT adapter. Do not cite as an adapter result.

---

## EXPERIMENT 4: NEEDLE-IN-HAYSTACK (CORRECTED 2026-06-26 — old 0% was a harness artefact)

Setup (corrected harness):
- Context lengths: 1,024 / 2,048 / 4,096 tokens
- Depths: 0.1 / 0.5 / 0.9 (position of needle in haystack)
- Samples: 5 per (context_len, depth, variant)
- Variants: base model (TinyLlama-1.1B-Chat), AwareLiquid M1 adapter (checkpoint 003000)
- Harness: bench_needle_m1_faithful.py — uses the instruct chat template AND rebuilds the
  faithful phase-5b model (MT adapter + PEFT LoRA), with an honest guard that aborts unless
  every adapter tensor maps onto the graph (verified 374/374 tensors, unexpected=0)

RETRACTION: The earlier "0% accuracy across ALL conditions" was INVALID. It came from a harness
that concatenated raw filler/needle/question tokens WITHOUT the instruct chat template — a format
that always scores 0.0 on an instruct-tuned base (documented in NEEDLE_FIX.md). It measured the
broken harness, not the model.

Corrected results (exact match, averaged over depth):
- Base    : 1024 = 0.867 | 2048 = 1.000 | 4096 = 0.000
- M1 adapter: 1024 = 1.000 | 2048 = 1.000 | 4096 = 0.000

Honest reading:
1. Within TinyLlama-1.1B's 2048-token window, retrieval is near-perfect for both base and adapter.
2. The adapter is at parity with the base in-window (slightly better at 1024 where the base misses
   one mid-depth needle), but with only 5 samples/cell and the base near saturation, a measurable
   retrieval UPLIFT remains INCONCLUSIVE — we report parity, not a win.
3. At 4096 tokens both collapse to 0.000 because that exceeds TinyLlama's 2048 RoPE training window
   — a genuine base-model limit (the frozen base never had a larger window), NOT an adapter failure.

Verdict: RETRIEVAL WORKS WITHIN THE BASE WINDOW. Extending beyond 2048 / proving an adapter uplift
needs a larger-window base and more samples — logged as future work.

---

## EXPERIMENT 5: HEBBIAN PLASTICITY ABLATION (M1 60.7M)

Setup:
- Hardware: CPU (reproducible)
- Model: AwareLiquid M1 60.7M, only Hebbian term active (all other liquid modules disabled)
- Hebbian LR sweep: off / 1e-4 / 1e-2 / 1e-1
- Seeds: 2 per setting, 150 steps each
- Dataset: WikiText-103

Results:
| hebbian_lr | PPL mean | PPL std | ΔPPL vs off | Gradient share |
|------------|----------|---------|-------------|----------------|
| off        | 336.9    | 6.6     | —           | 0              |
| 1e-4       | 337.2    | 7.2     | +0.32       | 8.3e-8         |
| 1e-2       | 337.2    | 7.2     | +0.32       | 8.3e-6         |
| 1e-1       | 337.2    | 7.3     | +0.35       | 8.5e-5         |

ΔPPL changes are within seed noise (±6–7 PPL). Hebbian gradient contribution is 4–5 orders of magnitude below main gradient. Verdict: INERT at current scale.

---

## SYSTEM BENCHMARKS

Memory: State-only streaming uses constant 4 KB regardless of context length. KV-cache streaming uses O(n): 108 KB at 1K tokens → 1,045 KB at 10K tokens.

Sparse resonance (k=2 vs dense k=5): +24% throughput with <0.34% mean logit divergence vs dense.

---

## INTELLECTUAL LINEAGE

Primary: Hasani et al. "Liquid Time-constant Networks" (MIT CSAIL / IST Austria, NeurIPS 2021). AwareLiquid's LTC core directly implements their ODE formulation.

Related work:
- Kirkpatrick et al. "Overcoming catastrophic forgetting in neural networks" (EWC, 2017) — AwareLiquid's consolidation mechanism
- Baars (1988), Dehaene et al. (2011) — Global Workspace Theory → GWT-B module
- Friston (2005) — Free Energy Principle → predictive coding module
- Hebb (1949) — "Neurons that fire together wire together" → Hebbian plasticity module
- Liquid AI LFM / LFM2 — same MIT lineage, different research axis

---

## CITATION

BibTeX:
@misc{awareliquid2026,
  title        = {{AwareLiquid} {M1}: Bio-inspired Continual Learning with
                  Global Workspace, Sleep Consolidation, and {EWC}},
  author       = {{AwareLiquid Project}},
  year         = {2026},
  howpublished = {\url{https://awareliquid.ai}},
  note         = {M-Series cognitive engine. Open research prototype.
                  Code: \url{https://github.com/everest-an/M1}}
}

APA:
AwareLiquid Project. (2026). AwareLiquid M1: Bio-inspired Continual Learning with Global Workspace, Sleep Consolidation, and EWC [Open research prototype, M-Series]. Retrieved from https://awareliquid.ai. Code: https://github.com/everest-an/M1

---

## FREQUENTLY ASKED QUESTIONS

Q: What is the AwareLiquid M-Series / O-Series distinction?
A: M-Series (flagship: M1) is the cognitive slow-thinking line — it layers GWT competitive broadcast, NREM sleep consolidation, and EWC continual learning on the LTC backbone. O-Series (flagship: O1) is the edge fast-thinking line — pure continuous-time liquid ODE, constant-memory state, no KV-cache, targets millisecond CPU latency. Both share the MT-LNN backbone; they differ in which cognitive modules are active and the deployment target.

Q: Is AwareLiquid production-ready?
A: No. It is an open research framework. O1 (48M from scratch) produces coherent short stories only. M1 (TinyLlama-1.1B + MT adapter) improves perplexity but generates repetitive outputs at higher temperatures. Neither is a general-purpose assistant.

Q: How is AwareLiquid different from Liquid AI?
A: Liquid AI (LFM) optimises frozen-weight models for maximum inference efficiency — targets industrial time-series, edge signal processing, autonomous driving. AwareLiquid builds cognitive architecture on top of the liquid backbone: GWT awareness, sleep-based memory consolidation, EWC online weight updates. Same academic lineage (MIT LNN), entirely different research axis. AwareLiquid is not affiliated with Liquid AI, Inc.

Q: What is catastrophic forgetting?
A: When a neural network trains sequentially on domain A then B, its performance on A collapses. In AwareLiquid M1 experiments: 244.9 → 10,806.9 PPL for dense baseline (×44 degradation). EWC reduces this to 1,878 PPL (×7.6 degradation). Threshold for "usable" is ≤5× — not yet achieved. Reported honestly as Early Signal.

Q: Does the needle-in-haystack benchmark work?
A: Yes, within the base model's context window. Re-tested 2026-06-26 with the instruct chat template (the earlier "0%" was a harness artefact from concatenating raw tokens without the template). Within TinyLlama-1.1B's 2048-token window retrieval is near-perfect: base 0.867/1.000 at 1K/2K, adapter 1.000/1.000. At 4096 tokens both drop to 0.000 — that exceeds the base's 2048 RoPE window, a base limit not an adapter failure. Adapter at parity in-window; a measurable uplift stays inconclusive at 5 samples/cell. Extending long-context retrieval is logged as future work.

Q: What is the API?
A: GET /v1/model (model info), POST /v1/completions (batch), POST /v1/completions/stream (SSE), POST /adapter/v1/completions/stream (M1 adapter SSE). Research-grade, no SLAs. See https://github.com/everest-an/M1 for local setup.