ARMS & HAT Memory Lab

About This Lab

The ARMS & HAT Memory Lab provides interactive tools to explore our breakthrough research in AI memory systems. These experiments demonstrate how:

• HAT organizes attention states into navigable hierarchical trees
• ARMS stores states at coordinate positions in 4096-dimensional space
• Memory retrieval achieves 100% recall with O(log n) complexity

The Core Innovations

HAT: Structure Over Learning

Traditional vector databases (HNSW, Annoy, FAISS) learn topology from data. HAT exploits known structure:

Traditional:  Points → Learn topology → Navigate
HAT:          Points → Use known hierarchy → Navigate directly

Result: 100% recall vs 70% for HNSW
        70× faster construction

ARMS: Position Is Memory

Traditional memory systems project states to lower dimensions, losing information at each step. ARMS stores states at their actual coordinate positions:

Traditional:  State → Project → Index → Retrieve → Reconstruct (lossy)
ARMS:         State → Store AT coords → Retrieve → Inject (lossless)

Result: 5,372× compression with exact restoration

Interactive Experiments

1. HAT Tree Visualization

Explore the hierarchical structure of a HAT index:

• Global root containing all memory
• Session nodes (conversation boundaries)
• Document nodes (topic groupings)
• Chunk leaves (individual attention states)

See how centroids propagate up the tree and how beam search navigates down.

2. 4096D Coordinate Explorer

Visualize how attention states cluster in high-dimensional space:

• t-SNE projections showing topic separation
• UMAP embeddings revealing structure
• Cross-topic similarity: -0.33 (excellent discrimination)

3. Memory Retrieval Demo

Watch HAT find memories in real-time:

1. Enter a query
2. See beam search expand candidates at each level
3. Observe centroid similarity scores
4. View retrieved attention states

4. Compression Analyzer (Coming Soon)

Analyze why attention patterns are highly compressible:

• ~90% of weights have less than 1% magnitude
• Similar queries → similar patterns (cacheable)
• Compression potential: 5,000-18,000×

Key Metrics

The Hippocampus Model

Our architecture mirrors human memory:

Technical Stack

• Core: Rust (performance-critical paths)
• Bindings: PyO3 (Python integration)
• Index: Custom HAT + FAISS baseline
• Encoder: Sentence Transformers
• Visualization: React Three Fiber + D3

Getting Started

from arms_hat import HatIndex

# Create index
index = HatIndex.cosine(1536)

# Start a conversation (session)
index.new_session()

# Add messages
for message in conversation:
    embedding = encoder.encode(message)
    index.add(embedding, message)

# Query memory
query = encoder.encode("What did we discuss about X?")
results = index.near(query, k=10)

# Retrieve with 100% accuracy
for state_id, similarity in results:
    print(f"Found: {state_id} @ {similarity:.3f}")

Research Applications

• Persistent conversation memory - Cross-session context
• Knowledge graph construction - Structured fact extraction
• Model debugging - Inspect attention patterns
• Compute caching - Skip redundant attention computation
• Multi-agent memory - Shared attention manifolds

This lab is part of ongoing research at Automate Capture. Experiments are interactive and run in your browser where possible.