A $50,000 factory robot resumes from exact state after power loss.The most advanced AI on the planet forgets everything when you close the tab.
We built the fix.
No LLM in the loop. No cloud. Zero cost per query. The system extracts deterministically and knows the exact boundary of what it has. Your data never leaves your machine.
Talk to ChatGPT. Switch to Claude. It already knows.
Switch to Cursor. It already knows.
Come back next week. It still knows.
One local server. Every AI tool shares the same understanding of you. Across sessions. Across providers. Across time.
Other memory products store what you said. Kenotic stores what the AI understood about you while you were saying it.
Two categories get called “memory.” They are not the same thing.
Type 1 · Retrieval memory
What you said.
Stores facts. “User lives in Michigan. Prefers Python. Has a sister named Mia.” When asked, it looks up the fact and injects it into the prompt.
Mem0, Rewind, Personal.ai, ChatGPT Memory, Letta, Zep.
Type 2 · Understanding continuity
What the AI understood about you.
Stores the AI’s evolving understanding of you. Not the words you said, but the thread you were on, the things left unresolved, the working model it was building. Switch to a new AI and it picks up where the last one left off.
Kenotic.
Why the labs with the capital to build this are incentivized not to.
OpenAI will never build something that makes Claude better. Anthropic will never build something that makes GPT better. A cross-host memory layer means helping your competitor. Their business models will not allow it.
The neutral layer has to come from outside.
What we built.
Deterministic extraction. No LLM in the loop. Everything here is verifiable.
Deterministic extraction
4,200 lines of grammar rules and spaCy pipelines. No LLM decides what to store. The system knows the exact boundary of what it has.
Five structured traces
Every utterance gets decomposed into five trace types: episodic, emotional, temporal, relational, and schematic. All stored as a single row.
Supersession
“I work at Netflix” gets stored. “I got a job at Google” marks Netflix as historical, Google as current. No LLM deciding what to update.
Cross-host via MCP
Works with Claude, GPT, Gemini, Cursor, and everything MCP-compatible. One local server. Install once.
Zero marginal cost
spaCy + SQLite + grammar rules. No cloud. No GPU. No per-query cost. Runs on any laptop.
The system that says “I don’t know”
Ask about something that was never stored. Every other system gives you a confident answer anyway. Kenotic abstains, because deterministic extraction knows what is and is not there.
The work, in public.
Three arXiv papers, two patent applications, and a working demo. All public. All before raising a dollar.
One primitive. Many applications.
Anything that has to act in context, against history, without being told what to do step by step. They all need the same layer underneath.
Industrial systems
A manufacturing line that carries forward a failure pattern from three months ago, without operator intervention, and pauses before the bearing fails again.
Healthcare devices
An insulin pump that decides the next dose by reconstructing the patient’s current situation, not by querying a database of past readings.
Robotics
A robot that does not need to be re-taught when the environment changes. The differences are updates to a state the robot is already holding.
Autonomous agents
An agent inside a codebase that does not start every session from scratch. The state of the project is already in the layer before the session begins.
The continuity layer is being designed by someone who has shipped production hardware.
Most AI infrastructure is designed by people who have never had to keep something running through a power cycle. The continuity problem is invisible from inside a Jupyter notebook. From inside a factory, it is the first thing you solve.
The roadmap ends in silicon. A continuity node that any device can integrate. The path there starts on the floor of a manufacturing line.
Three years designing industrial automation systems. $1.5M of production machinery shipped to Fortune 500 manufacturers. The machines build iPhone bodies and Tesla components. They run 24/7. Downtime is not an option.
Industrial controllers hold persistent state across power cycles, restarts, and shift changes. They are deterministic by construction. The continuity layer is the same idea, one stack higher.
Mechatronics. Electronics and Computer Engineering. Information Systems (M.S., Central Michigan University). Silicon, board, firmware, software, and the manufacturing process underneath all of it.
Read the full thesis.
Thirteen sections. Roughly 8,000 words. Why continuity is the missing layer, why the scaling wall makes it necessary now, and what the architecture looks like.
Publications
Software serves people.
Your data is yours.
It stays on your device. Privacy is physics, not policy. We do not collect what we cannot see.
You are not the product.
You are not training data. You are not an engagement metric. The system carries your life forward, not extracts value from it.
AI should be fair.
It works for everyone. Not just the users who generate the most revenue.
We publish our evidence.
Our results include our failures. If we cannot prove it, we do not claim it.
Built from the ground up. Shipping as infrastructure.
Kenotic Labs is an AI infrastructure company building the continuity layer for AI. Infrastructure that preserves and reconstructs the living state of a situation across time. The core engine works today: deterministic extraction, cross-host memory via MCP, and a working demo that carries understanding across Claude, GPT, Cursor, and anything MCP-compatible. Founded in 2025 by Samuel Sameer Tanguturi. Based in Michigan, USA.
Three arXiv papers, two patent applications, an open benchmark on GitHub, and a narrative corpus on Hugging Face. All published, all verifiable, all before raising a dollar.
Let's talk.
Partnership inquiries, investment, research collaboration, or joining the team.
Join the waitlist.
Be first to access the Continuity SDK or try Raya when it launches.