Skip to content
All products

Nema Cortex

Nema Cortex is a small language and memory system that we built from scratch, with no transformer. It runs on a single CPU core, keeps reading text, and answers questions about what it has learned. When it does not know something, it says so instead of inventing an answer. It is a research preview and not an assistant.

Status
Research preview, in internal testing
Runs
on one CPU core, no graphics card, on our own server
Response time
about 50 milliseconds per question
Learns from chat
never

Ask it yourself

This is the real instance, not a video. It learns nothing from your questions, and your text is not stored. Each question may have up to 300 characters, and you can ask ten per minute. Cortex speaks German.

Try one of these questions
It does not know this one

What it is

Cortex has three parts: a model that reads text byte by byte and predicts it, a search index over dialogues and Wikipedia text, and a concept layer that extracts facts such as “Rome is the capital of Italy” from Wikipedia articles and chains several of them. Everything runs in a single process on one core. It keeps reading its corpus (the German Wikipedia corpus is about 1.5 GB), saves its state, and answers through a simple interface. It learns without backpropagation, the method used to train large language models.

Where an answer comes from

Under every answer in the demo you see where it came from. That is the honest side of Cortex: you can tell whether it knew something or had nothing.

From memory
The question matches a stored question well. Cortex returns the stored answer verbatim.
From the concept store
Cortex extracted facts from Wikipedia and chains them, for example “Rome is the capital of Italy, so Rome lies in Italy”. The answer comes with the sentence the fact was taken from.
I do not know
Nothing matches well enough. Cortex then gives a fixed “Darüber weiß ich nichts.” and does not guess.

What it is good for

Quick factual questions about what is in the Wikipedia texts it has read, for example capitals, where places are, birthplaces and simple definitions, on tiny hardware and with an honest no when it does not know. It is also our research subject, to understand how far you get without a large language model.

What it is not

Cortex is not an assistant. It does not follow instructions, write longer texts, reason over several steps or use tools. It does not hold a conversation either: every question is answered on its own. It looks up what it has learned and repeats it without understanding it. On 200 dialogue questions it had never seen, 45.4 % of the answer bytes came verbatim from the retrieved answer, but only 5.6 % of the rare words of the true answer. Asked for the square root of 36, it answered “64” in one test: right form, wrong number. And it only knows what is in its sources. It is far from the large models that our agents in Nema Crew use, and it does not replace them.

How it differs from large models

It learns continuously without retraining
Cortex simply keeps reading. Facts we tell it can be repeated at once, without a training run. It learns nothing from the questions in the demo.
Tiny footprint
One CPU core, no graphics card, about 2 GB of memory. An answer takes around 50 milliseconds.
An honest no
Where large models often invent a smooth answer, Cortex says it does not know.
Traceable
Every answer names its source, and answers from the concept store come with the sentence the fact was taken from.

Numbers

All numbers come from the service's documentation and our research record. They are our own measurements on small samples, not an independent test.

52 ms
median time for an answer, 95 percent of answers came within 54.6 ms. Measured with 40 requests on a test instance with a small text corpus, not with the full corpus.
1 core
The whole service runs on one CPU core, using about 2.1 GiB of memory in the test. With the full corpus we expect about 2.7 GB.
100 %
of nonsense questions (“Was ist ein Blurfnak?”) were answered honestly with “I do not know”. For questions about articles that are not in the index it was 59.0 %.
100 %
of questions about articles that are in the index were answered (measured on 3 September 2026, service set to “I do not know” for weak matches).

Sources: README.md (section “What it can and cannot do”) and docs/OPERATIONS.md (section “Measured on srv2”) in the repository nema/models/cortex, which cite the research measurements of 2 and 3 September 2026.

Where it has limits

  • It speaks German only and knows only what is in its sources, mainly the German Wikipedia corpus and dialogue files.
  • It keeps nothing from one question to the next. “And how many people live there?” is not understood.
  • The concept layer knows only a few kinds of question: what something is, capital, location, birthplace. Everything else falls back to memory or to “I do not know”.
  • The facts come from Wikipedia introductions and are extracted automatically. They can have gaps and errors.
  • It is an internal research system. Customer data does not belong in it.