Memory that doesn't forget.

Norax AI is an autonomous agent with a 10-layer cognitive memory architecture. It runs on local GPUs, never ships your data to a third-party model, and scored 100% on the hardest long-term memory benchmark in AI research — using a free open-source model.

LongMemEval Oracle · 48/48 $0 inference cost 100% local ICLR 2025 benchmark

The problem

Most AI assistants forget you the moment you close the tab. They lose context, repeat themselves, and treat every conversation as if it were the first. Enterprise teams pay tens of thousands per month for systems that still can't remember a user preference across a weekend.

What we built

A memory-first AI agent. Ten dedicated memory layers — episodic, semantic, procedural, working, rolling summary, daily logs, knowledge graph, scratchpad, cache, and intel — retrieved in parallel in 7–9 seconds. A 2,200:1 compression ratio turns 167KB of context into the 76 bytes that actually matter for your query.

It runs on a dual-GPU workstation we built ourselves. Inference is free because the model is open source. Memory retrieval is free because we host the embedding model locally.

Numbers

48/48LongMemEval Oracle
100%accuracy · prev SOTA 91.4%
$0per-message inference cost
10parallel memory layers
2200:1context compression
7–9sfull-memory retrieval

Who we are

Norax AI is a solo-founded, bootstrapped AI research and product company. We build on OpenClaw, an open-source agent runtime, and contribute improvements upstream. Our mission: agent AI that remembers, reasons, and works while you sleep — without renting intelligence from a frontier lab.

What's next

Get in touch

Investors, press, partners: [email protected].

Try the demo Read the paper