LlamaIndex Hyperspace Integration: Spatial AI Memory Infrastructure

Building the Episodic Memory for the AGI Era.

This is the official LlamaIndex integration for HyperspaceDB — the world's first Spatial AI Engine. It models information exactly how the physical world and human cognition are structured: as hierarchical, spatial, and dynamic graphs.

🧠 Why Spatial AI for LlamaIndex? (Beyond RAG)

Traditional vector databases were built to search static PDF files for chatbots. HyperspaceDB provides the primitives for autonomous agents and robotics:

• Fractal Knowledge Graphs: Euclidean vectors fail at hierarchies. Our Poincaré & Lorentz models compress massive trees (like codebases or medical taxonomies) into low-dimensional spaces, reducing RAM usage by 50x without losing semantic context.
• Continuous Reconsolidation: AI agents need to "sleep" and organize memories. With our Fast Upsert Path and Riemannian Math SDK (Fréchet mean, parallel transport), your indexers can continuously shift and prune vectors dynamically.
• Heterogeneous Tribunal Framework: Natively support the confrontational model of LLM routing (Architect vs. Tribunal) directly on the vector graph. Calculate a Geometric Trust Score to verify logical path lengths and detect hallucinations.
• Edge-to-Cloud Delta Sync: Drones and humanoid robots can't wait for cloud latency. HyperspaceDB runs directly on Edge hardware, using Merkle Tree Delta Sync to asynchronously handshake and sync memory chunks with the Cloud.

📦 Installation

pip install llama-index-vector-stores-hyperspace hyperspacedb

🛠 Usage

Hyperbolic Memory Initialization

from llama_index.vector_stores.hyperspace import HyperspaceVectorStore
from llama_index.core import StorageContext, VectorStoreIndex
from hyperspace import HyperspaceClient

client = HyperspaceClient("localhost:50051", "API_KEY")

vector_store = HyperspaceVectorStore(
    client=client,
    collection_name="agent_memory",
    metric="lorentz",  # Use hyperbolic geometry for complex hierarchies
    dimension=64
)

Hallucination Detection (Tribunal Framework)

Evaluate the structural trust of an LLM claim by verifying the logical path length between concepts in latent hyperbolic space:

from hyperspace.agents import TribunalContext

# 1.0 = Truth (Identical), 0.0 = Hallucination (Disconnected)
score = client.evaluate_claim(concept_a_id=12, concept_b_id=45)
print(f"Geometric Trust Score: {score}")

Multi-Geometry Spatial Filters

Prune search results by geometric regions:

from llama_index.core.vector_stores import MetadataFilters

filters = MetadataFilters(
    filters=[
        # Spatial Sphere (Ball) Pruning
        {"key": "location", "value": {
            "$in_ball": {"center": [0,0,0, ...], "radius": 0.5}
        }}
    ]
)

Native Vector-Graph (KAG) & Topology Linking

Connect document nodes and entity concepts directly in the in-engine graph:

# Explicit directed linking
vector_store.link(source_id=1, target_id=2, relation="OWNS", weight=1.0)

# Batch linking
vector_store.link_batch([
    {"source_id": 1, "target_id": 2, "relation": "OWNS", "weight": 1.0},
    {"source_id": 2, "target_id": 3, "relation": "DEPENDS_ON", "weight": 0.95},
])

# Unlink edge
vector_store.unlink(source_id=1, target_id=2, relation="OWNS")

# Declarative edge induction from document metadata
created = vector_store.induce_edges_from_metadata([
    {"field": "file_path", "rule_type": "path_hierarchy", "relation": "CONTAINS"},
    {"field": "parent_id", "rule_type": "foreign_key", "relation": "OWNS"},
])
print(f"Induced {created} explicit edges directly from metadata")

⚡ Performance: Reflex-Level Speed

Built on Nightly Rust. Our ArcSwap Lock-Free architecture and SIMD f32 intrinsics deliver up to 12,000 Search QPS and 60,000 Ingest QPS for real-time robotic memory.

📖 Documentation

• The Case for Hyperbolic Memory
• HyperspaceDB Official Documentation

📄 License

Apache-2.0. Copyright © 2026 YARlabs.