---
language: en
#license: CC-NC
tags:
- vector-symbolic-architecture
- holographic-reduced-representations
- sparse-distributed-memory
- semantic-hashing
- cognitive-architecture
- memory-systems
- no-gpu
library_name: numpy
pipeline_tag: feature-extraction
---

# Holographic Neural Mesh (HNM) v3.0

A **deterministic sparse semantic substrate** for cognitive systems. No GPU required.

## Model Description

HNM is **not** a language model or embedding model replacement. It is a **cognitive memory layer** providing:

- ✅ Constant-time encoding (O(1) regardless of corpus size)
- ✅ 99% structural sparsity (102× FLOPS reduction)
- ✅ Deterministic representations (same input → same output → always)
- ✅ Semantic discrimination (negation, role reversal, synonyms)
- ✅ Associative binding/unbinding (key-value memory)
- ✅ Pure NumPy (no GPU, no PyTorch, no TensorFlow)

### What HNM IS

| ✅ HNM Is | ❌ HNM Is Not |
|-----------|---------------|
| Semantic memory substrate | Language model |
| Symbolic binding engine | Next-token predictor |
| Deterministic cognitive layer | Embedding model replacement |
| Associative recall system | Foundation model |

## Intended Uses

- **Agent memory backbone**: Long-term memory for LLM agents
- **Semantic routing**: Content-addressable dispatch
- **Variable binding**: Compositional reasoning substrate
- **Edge deployment**: Cognitive processing without GPU
- **Distributed consensus**: Deterministic state for multi-agent systems

## How to Use

```python
from hnm_v3 import HolographicNeuralMeshV3, HNMConfig

# Initialize
hnm = HolographicNeuralMeshV3(HNMConfig())

# Encode text
pattern, stats = hnm.forward("Machine learning is fascinating")
print(f"Latency: {stats['inference_time_ms']:.2f}ms, Sparsity: {1-stats['active_ratio']:.1%}")

# Semantic similarity
sim = hnm.similarity("I am happy", "I feel joyful")  # ~0.87
sim = hnm.similarity("dog bites man", "man bites dog")  # ~0.52 (role reversal detected)

# Memory storage and retrieval
hnm.encode_and_store("Deep learning uses neural networks")
hnm.encode_and_store("The stock market crashed today")
results = hnm.search("Tell me about neural networks", top_k=3)

# Associative binding
bound = hnm.bind("capital of France", "Paris")
recovered = hnm.unbind(bound, "capital of France")  # ≈ Paris vector
```

## Benchmarks

### Semantic Discrimination (7/7 Pass)

| Test | Pair | Score | Target | Status |
|------|------|-------|--------|--------|
| Negation | "alive" / "not alive" | 0.48 | < 0.50 | ✅ |
| Role Reversal | "dog bites man" / "man bites dog" | 0.52 | < 0.70 | ✅ |
| Paraphrase | "happy" / "joyful" | 0.87 | > 0.70 | ✅ |
| Unrelated | "neural networks" / "fishing" | 0.01 | < 0.30 | ✅ |

### Scaling (Constant Time)

| Corpus Size | TF-IDF | BM25 | HNM |
|-------------|--------|------|-----|
| 20 docs | 0.03ms | 0.04ms | 1.78ms |
| 2,000 docs | 2.45ms | 3.60ms | **1.62ms** |
| **100× growth** | **78× slower** | **98× slower** | **0.9× slower** |

### Resource Efficiency

| Metric | Value |
|--------|-------|
| Sparsity | 99% |
| FLOPS reduction | 102× |
| Inference latency | ~3.5ms |
| GPU required | **No** |

## Architecture

```
Input → Semantic Encoder → Holographic Projection → Interference Layers (×8) → Memory
            ↓                      ↓                        ↓                    ↓
      Word vectors +         Complex pattern          FFT + Phase mixing    Cleanup memory +
      Negation handling      Phase = semantics        99% sparsification    Iterative decoding
```

**Key Components:**
- **Dual-channel encoding**: Semantic (meaning) + Structural (order)
- **Circular convolution binding**: `bind(key, value)` → `unbind(bound, key) ≈ value`
- **Cleanup memory**: Iterative extraction from superposition
- **Hierarchical storage**: 16 slots with saturation monitoring

## Limitations

1. **Semantic priors are hand-crafted**: ~40 word clusters cover common vocabulary but not open-domain language
2. **Similarity is ordinal, not metric**: Scores are internally consistent but not calibrated to SBERT
3. **This is a substrate, not a system**: Provides memory/binding, not language generation

## Citation

```bibtex
@software{stone2024hnm,
  author = {Stone, Kent},
  title = {Holographic Neural Mesh: A Deterministic Sparse Semantic Substrate},
  year = {2024},
  publisher = {JARVIS Cognitive Systems},
  url = {https://huggingface.co/jarvis-cognitive/hnm-v3}
}
```

## Lineage

HNM builds on established cognitive architecture research:
- **Holographic Reduced Representations** (Plate, 1995)
- **Sparse Distributed Memory** (Kanerva, 1988)
- **Hyperdimensional Computing** (Kanerva, 2009)
- **Vector Symbolic Architectures** (Kleyko et al., 2023)


## Contact

Kent Stone - JARVIS Cognitive Systems