Tzivia-AI

Tzivia

Middleware that makes LLMs faster and cheaper.

Tzivia is a persistent memory and reasoning layer that attaches to any large language model. Instead of re-feeding entire context windows on every call, Tzivia retrieves only the relevant knowledge chains — cutting search space, latency, and token cost by orders of magnitude.

⚡ 3,136× faster routing than linear search at 100k items 🧠 88ms to retrieve 25 knowledge chains from 1,000,000 tokens 🗜️ 99.6% search-space reduction (1M tokens → 3,861 synthesis nodes)

The Problem

Every LLM call today pays the same tax: context has to be stuffed into the prompt, tokenized, and attended over from scratch. As knowledge bases grow, this gets slower and more expensive linearly — or worse. Retrieval-augmented approaches help, but most treat memory as a flat vector store with no structure, no reasoning, and no persistence of insight.

The Approach

Tzivia models memory as a typed, decaying knowledge graph with a dedicated routing algorithm on top:

• ARIA routing algorithm — sub-linear traversal that reaches relevant nodes without scanning the full store. Benchmarked at 3,136× faster than linear search at 100k items.
• 4D vector scoring (Vec4) — every node is scored across four independent dimensions: frequency, accuracy, connectivity, and inference weight. Retrieval balances all four rather than relying on similarity alone.
• Typed semantic relationships — 10 distinct edge types (not just "related to"), enabling a formal logic engine to reason over the graph with defined inference rules.
• Synthesis nodes — a persistent insight layer that doesn't decay. Discovered patterns are promoted into durable knowledge instead of being recomputed every session.
• MCP server architecture — Tzivia attaches to LLMs through the Model Context Protocol, making it a drop-in memory layer for any compatible model.

Benchmarks

Methodology and reproduction details available on request.

Applied Result: DNA Decay Finder

To validate the graph-decay mechanics on real data, Tzivia's decay model was applied to cancer genomics. The DNA decay finder detects mutation hotspots and was validated against the public TCGA cancer genomics dataset — demonstrating that the same routing and decay machinery that accelerates LLM memory generalizes to structured scientific data.

Architecture at a Glance

        ┌─────────────┐  LLM ──│ MCP Server  │── attachment point        └──────┬──────┘               │        ┌──────▼───────────────────────────┐        │  ARIA Routing  (sub-linear)       │        ├──────────────────────────────────┤        │  Typed Knowledge Graph            │        │   • Vec4-scored nodes             │        │   • 10 typed relationship edges   │        │   • Logic engine (formal rules)   │        ├──────────────────────────────────┤        │  Synthesis Layer  (no decay)      │        └──────────────────────────────────┘               │        ┌──────▼──────┐        │   SQLite    │  local persistence        └─────────────┘

• SQLite persistence — zero-configuration local storage; the full graph lives in a single file
• Single-tenant by design — your memory graph stays on your machine

Roadmap

• PostgreSQL / Supabase backend for hosted, multi-tenant deployments
• Zero-knowledge encryption for multi-tenant mode
• Real-time 3D knowledge-graph visualization (Fibonacci-sphere layout, force-directed)

Status

Tzivia is in active development at v5.1. Core components — ARIA routing, Vec4 scoring, the typed edge store, the logic engine, and the validated DNA decay finder — are working and benchmarked. Public launch targeted for December 2026.

About

Built by Matthew McBerry, founder & CEO of Dorcas Innovations.

The full engine is under active development and kept private as core IP. This repository is a public overview — for a live demo, benchmark reproduction, or partnership inquiries, reach out.

📧 [email protected]

Tzivia — persistent memory and reasoning for any LLM.