Qriton HLM

An editable, auditable memory layer for AI systems.

Qriton HLM is an editable, auditable memory layer for AI systems: certified edits to what a model knows, an explicit memory that grows without retraining, and a record of which stored items produced each answer, on a host model that stays fixed.

A visualisation of an HLM energy landscape with labelled basins

Why the memory sits beside the model.

Knowledge held in a model's weights can only be changed by retraining, and nobody can say afterwards what else changed. HLM keeps the host model fixed and holds knowledge in an explicit memory beside it, so each record can be added, corrected or removed on its own, with a certificate for the change.

The host can be one of our reference checkpoints or another strong model.

What it does.

Certified edits

Energy Language applies one change, checks that it reached its target and left held-out behaviour unchanged, and keeps it only if the check passes. Any change can be reverted exactly.

Growing memory

An explicit store beside a frozen host model, tested to one million records with a bounded number of candidates per query.

Record per answer

Each answer records which stored items it used, so any exchange can be replayed.

How it works.

01

Define the task

Agree the decision, the data and the measures of success with your team.

02

Choose the host

Use one of our reference checkpoints or your own model, and measure it against standard models on your data.

03

Add the memory

Load your records, apply corrections with their certificates, and keep the log for review.

50,000

GPU hours on Leonardo, through the EuroHPC AI Factory

At a glance.

Reference models
HLM5 for language; HLM-Vision, HLM-Spatial and HLM-Audio; HLM-Micro for activity, keyword and gesture recognition. Research models, used as hosts for the memory layer.
Inputs
Text, documents, images, audio, 3D and sensor streams.
Deployment
Your servers, your private cloud or air-gapped sites.
Tools
Energy Language, published as the qriton-hlm package: certified changes that can be reverted exactly.
Access
Controlled early access for approved teams. Several research models, including HLM5-3B Base, are public on Hugging Face.

Questions.

Is HLM a better language model?
No. Our reference checkpoints are research models and we do not claim that they outperform standard models of similar size. HLM is the memory layer around a host model: certified edits, a growing explicit memory and a record per answer.
Can we run HLM on our own hardware?
Yes, on your own servers, including air-gapped sites.
How do we get access?
HLM is in controlled early access for approved teams. HLM5-3B Base is public on Hugging Face as an inference-only research release.

Bring us the decision that has to be right.

BriefingPilot on your dataProduction