hyperspace-memory (Python) 🚀

Drop-in replacement for Mem0, Zep, and MemGPT powered by the HyperspaceDB native memory engine with 1-bit ADC quantization, MRL in-RAM cascades, and Lorentz hyperbolic spaces.

⚡ Key Advantages over Mem0 & Zep

• 100x Lower Search Latency: Sub-millisecond (< 0.5 ms) retrieval via MRL in-RAM cascade indexing.
• 98% RAM & Storage Reduction: Native extreme 1-bit ADC and turbo 4-bit Lloyd-Max quantization (compressed down to ~100B per vector).
• Zero Extra LLM Calls: Direct vector and graph traversal memory operations without forcing expensive LLM parsing for every insert.
• 100% Mem0 API Compatibility: Replace from mem0 import Memory with from hyperspace_memory import Memory without altering your application code.

📦 Installation

pip install hyperspace-memory

🛠 Quick Start (Mem0 Drop-In Migration)

from hyperspace_memory import Memory

# 1. Initialize Memory (connects to HyperspaceDB)
memory = Memory(config={
    "host": "the.yar.ink",
    "api_key": "YOUR_YARINK_API_KEY",
    "quantization": "medium_plus"  # "none" | "medium_plus" | "turbo" | "extreme"
})

# 2. Add Memory
res = memory.add(
    "User prefers Rust and dark mode UI over Python and light mode.",
    user_id="user_123",
    metadata={"category": "preferences"}
)
print("Added Memory:", res)

# 3. Search Memory
results = memory.search("What programming language does the user prefer?", user_id="user_123", limit=3)
for item in results:
    print(f"Memory: {item['memory']} (Score: {item['score']})")

# 4. Get All Memories for User
all_memories = memory.get_all(user_id="user_123")

# 5. Cognitive Consolidation (Fréchet Mean on Hyperboloid)
consolidation = memory.consolidate(
    "programming languages",
    summary_text="User strongly favors Rust and systems programming."
)
print("Consolidated concept:", consolidation["summary"])

# 6. Anti-Hallucination Claim Verification (Lorentz Geodesic Trust)
verification = memory.verify_claim(
    "User prefers Rust and dark mode UI over Python and light mode.",
    "User always requests Python notebooks with light themes."
)
print("Trust status:", verification["status"], "Score:", verification["trust_score"])

# 7. Delete & Reset
memory.delete(results[0]["id"])

🧠 Cognitive Operations

• memory.consolidate(topic_query, ...): Synthesizes a cluster of related episodic memories into a single abstract semantic concept using the Fréchet Mean on the hyperboloid.
• memory.verify_claim(premise, conclusion, threshold=0.30): Computes Geodesic Trust Score (Lorentz $\mathbb{H}^{32}$ + Cosine 96D) to verify that an LLM conclusion does not hallucinate.
• memory.link(source_id, target_id, relation="OWNS", weight=1.0): Links two memory nodes or concepts with an explicit directed edge in the in-engine CSR graph.
• memory.link_batch(edges): Atomic bulk edge insertion for rapid knowledge ingestion.
• memory.unlink(source_id, target_id, relation="OWNS"): Zero-lag deletion of graph edges.
• memory.induce_edges(rules): Declaratively builds edges from metadata attributes (path_hierarchy, foreign_key, exact_match).
• memory.list_sessions(limit=100): Discovers and enumerates active conversational sessions.
• memory.explore_hierarchy(concept_id, direction="down"): Navigates conceptual taxonomy trees in Lorentz space.
• memory.stats(): Returns vector counts, cascade configuration, and collection status.

📄 License

MIT © YARlabs.