# Cortex by SKYNETLAB

> Hosted semantic memory that any AI connects to over MCP. Every write passes a quality gate that rejects roughly 80% of attempts as redundant, contradictions are tracked instead of overwritten, and answers cite their sources.

Cortex is memory infrastructure, not a model. It lives alongside the assistants you already use (ChatGPT, Claude, or any MCP-capable client) and holds what they learn, so switching model does not mean starting from zero.

What makes it different is what it refuses. Most memory layers store everything and degrade: retrieval gets worse as near-duplicate notes pile up. Cortex classifies every write before persisting it, and one of four things happens:

- **Accepted**: the information is new and is stored.
- **Corroborated**: it is already known, so instead of creating a duplicate the existing memory's corroboration counter and trust score go up.
- **Contradicted**: it clashes with an existing memory; both are kept and the conflict is flagged as a first-class object, never silently overwritten.
- **Rejected**: it adds nothing. It is not stored, and it is not billed. This is the most common outcome, around 80% in production.

Rejection is synchronous and explained: the write call returns the verdict, the reason, and the ID and similarity of the nearest existing memory. A lightweight reject log (metadata only, no payloads, 30-day retention) is queryable so the rate can be audited.

Built by Filippo Pilotta, independent researcher and founder of SKYNETLAB (Bergamo, Italy). Patent pending, Italian Patent Office (UIBM) filing 102026000014026 of 15 May 2026; trademark 302026000119239.

## Product

- [Home](https://skynetlab-cortex.com/): what Cortex is, the quality gate in live numbers, and how it fits the AI tools you already use.
- [Use cases](https://skynetlab-cortex.com/casi-uso): concrete scenarios where a shared, model-independent memory changes the outcome.
- [Pricing and plans](https://skynetlab-cortex.com/abbonamenti): plans and the free trial.
- [Connect to Claude](https://skynetlab-cortex.com/claude): step-by-step MCP setup, roughly two minutes.
- [cortex-memory skill](https://skynetlab-cortex.com/skill): usage instructions to give your AI so it knows when to read from and write to Cortex.
- [FAQ](https://skynetlab-cortex.com/faq): common questions about setup, data handling, and behaviour.
- [Comparisons](https://skynetlab-cortex.com/confronti): Cortex versus Mem0, Zep and Graphiti, Letta, Supermemory, ChatGPT built-in memory, Claude built-in memory and do-it-yourself setups, with verifiable facts, sources and the cases where the alternative is the better choice.

## Evidence

- [Benchmark](https://skynetlab-cortex.com/benchmark): evaluation on LoCoMo (1,986 questions) with the full diary of measurement errors across four harness iterations, including the runs that produced numbers too good to be true. Strict J-score 55.8% with the gate active, 69.1% with full recall: the ~13-point gap is the cost of governance, quantified rather than hidden.
- [Academic paper](https://skynetlab-cortex.com/paper): the Coherence Metric work behind the system. DOI 10.5281/zenodo.17825268.
- [Patent](https://skynetlab-cortex.com/brevetto): filing details and scope.
- [Sustainability](https://skynetlab-cortex.com/sostenibilita): why refusing most writes means less compute and less energy: the saving is a consequence of the filter, not a separate feature.

## Optional

- [AI transparency](https://skynetlab-cortex.com/trasparenza-ai): how AI is used in building and running this product.
- [Contact](https://skynetlab-cortex.com/contatti): how to reach SKYNETLAB.
- [Privacy policy](https://skynetlab-cortex.com/privacy)
- [Legal notes](https://skynetlab-cortex.com/note-legali)