CORTEX Premium A licensed tier of the CORTEX memory layer
The commercial memory layer for agents that cannot afford to forget. Accuracy-tuned recall, supported for production.
Quartz is the accuracy-optimized build of the memory pipeline behind our open source CORTEX engine. Same cognitive architecture. Tuned, hardened, and benchmarked for teams that ship agents to real users.
Long-term memory is only useful if the right fact comes back at the right moment, across hundreds of sessions, weeks apart. That is what LongMemEval measures: can an agent answer a question whose evidence is buried deep in its history?
Every question surfaced its evidence in the top ten. The oracle split is now easy for modern retrieval, so we treat this as the floor Quartz starts from, not the result that matters.
Retrieval is half the job. Quartz adds the answering stage on top of it: observation extraction, a tuned answerer, and a retry pass when the model abstains, so the fact that was retrieved is the fact that gets used. When we publish end to end LLM-judge scores for that stage, they go on the bench with the harness and judge prompts, next to the retrieval scores, so you can rerun them.
CORTEX runs inside every agent ATERNA builds. Quartz is the same architecture, tuned and licensed for teams that build their own agents. Drop it in when you need maximum recall accuracy and a license you can build a business on.
| Tier | i.CORTEXin ATERNA agents | ii.Quartzlicensed tier |
|---|---|---|
| License | Included with every ATERNA agent | BSL 1.1, source-available, commercial |
| Best for | Businesses running an agent ATERNA built | Teams shipping their own agents to real users |
| Memory | Full hippocampal pipeline | The same pipeline, plus a tuned answering stage |
| Setup | Run by ATERNA for you | Drop-in provider, no rewrite |
| Support | Included in your plan | Direct from ATERNA |
Quartz sits behind the same MemoryProvider interface as CORTEX. In Meridian, our agent runtime, switching is one line. If the package is not present, the runtime falls back to CORTEX, so an agent always boots.
Your agent's safe-memory screen, the recall-stage defense that quarantines memory-poisoning attacks, works identically on every provider. Upgrading recall accuracy never weakens your security posture. This is the same switch ATERNA's own production agents run on, every turn.
# Default
MERIDIAN_MEMORY_PROVIDER=cortex
# Licensed tier, one flag
MERIDIAN_MEMORY_PROVIDER=quartzSix things a Quartz license includes, each one checkable before you sign.
BSL 1.1. Production use inside Meridian runtimes licensed by ATERNA. The license converts to Apache 2.0 on 1 May 2030, so the terms have a known end.
You read the code you are running. The answering stage, the observation store and the recall router ship as source, not as a black box behind an API.
Every licensed agent runs on its own CORTEX instance with its own keys. Nothing is pooled across customers, and that isolation is written into the license.
A charter of anchors for each agent and a consistency check over its actual output against them. Drift from stated values gets flagged, not discovered later.
Every release of the licensed pipeline for the term, including the tuning that moves the numbers on the bench.
Integration help, tuning on your corpus, and a roadmap you can influence. Direct, not a ticket queue.
We work directly with AI teams and technical founders who need this level of recall in production. Bring what you are building to the call and we get you running on Quartz.