Cortex
an AI-native streaming datastoreIf I only had a brain.
Between a model’s general intelligence and the intelligence an enterprise actually needs lies a gap of business context. Closing it is, at bottom, a data problem. The next generation of enterprise AI will not be limited by model intelligence alone. It will be limited by whether enterprises can turn their data into real-time, trusted, and actionable context.
Every contribution Cortex makes to its own state is recorded as something that can be withdrawn. When a fact is taken back, it does not become hidden or deprioritised. It stops being reachable at all.
At a glance
liveThe smartest thing you can do with Cortex is reduce token usage while saving money on skyrocketing AI inference costs.na8ve
Brain
replaying · structure heat
Reading the brain
The console draws Cortex as a brain because that is how it is built: nine structures, each doing one kind of work, connected by the tracts the work actually flows through.
- Heat is work. A structure glows in proportion to what it is doing, relative to this session's running peak - not to any capacity.
- Channels only through tracts. A structure with no channels still glows; a signal travels only along a drawn edge.
- Green excites, violet inhibits. Concept units are a separate scale: activation, never heat. Violet subtracts from what it charges.
- A strike is a discharge, not a packet in flight. A unit's place in its grid is layout; only a drawn edge is a relation.
Console
a recorded session · 10× speedWhat you are watching
Eleven minutes of the console, played at ten times speed. Questions arrive, memories are written and retrieved, the structures light as they work, and the trace underneath records every signal.
The console is not a dashboard bolted onto a database. It follows the datastore itself, reactive and streaming: it redraws as Cortex's state changes, instead of querying a report after the fact.
Structures
nine, each with one jobWhere observations arrive. Every write lands here first, as a sensory row, before anything decides what it means.
The mantle: live memories, the durable ones, and those that were retracted or went stale. The shell you see wrapping the brain.
The relay. It opens sessions, retrieves what a turn needs, and assembles the context a question is answered from.
Plans the answer. One plan per turn that retrieved something, written down where you can read it.
What leaves: the model's answer and what it cost to produce, token by token.
Selection and gating. The control signals that start, stop and route the work of the other structures.
Checks the work. It reads sources and evidence in a sealed sandbox and records what it found.
Folds many observations into a standing convention - the pattern behind the events, kept as its own memory.
Consolidation, like sleep: epochs that fold what was learned into the model's own weights.
Percentages are the frame shown in the brain view above.
Memory
continuous, not batchedA brain does not query itself with a "where is my memory" function. It notices, abstracts, and presents what it needs. Cortex keeps memory at three levels and moves it up as it earns its place:
- Sensory - what just arrived, exactly as it arrived.
- Episodic - what happened: events, with where they came from.
- Semantic - what is generally true: conventions the abstractor has folded out of many episodes.
It streams. Memory is updated as observations arrive, and consolidation folds what was learned into the model's own weights - so what it knows changes the way it thinks, continuously.
Retraction
withdrawn, not hiddenRetraction is at the heartof Cortex's memory system. Because every contribution is recorded as a change in state, any event can be taken back - and taking it back changes that state of the memory, in real-time.
- It cannot be retrieved - not by a search, not even by its own vector.
- What was built on it is marked. Memories derived from the fact are flagged stale, so nothing downstream keeps trusting it quietly.
- You can show the proof: the search that no longer finds it, at the moment it was withdrawn.
A retracted memory cannot be retrieved, even by its own vector - verified 16/16.
Audit
built to be watchedAn organisation evaluating an AI system need more than a certificate, it needs auditability and governance. It wants proof of how it actually thinks. Cortex is built for visibility, which is what makes it auditable - and auditability is the foundation governance stands on.
The institution
private, not personalisedCortex provides a AI model for your organisation, not a generic chatbot. Each organisation gets its own learning model with its own memory space. It is kept apart from every other: a search in one organisation's memory cannot draw a single candidate from another's.
Cortex is designed for organizations requiring uncompromising data privacy and cultural alignment. Your organizations model is trained continuously to your specific needs and business domain. This ensures your sensitive information remains secure while empowering your staff with AI tools authentically tailored to your unique institutional identity.
na8ve agent
a terminal agent with memoryA barebone, open-source agent for your terminal, with long-term memory. It connects to Cortex: it keeps what matters across sessions, answers why questions from what it holds, cites where each answer came from, and lets you withdraw any memory for good.
- Install. npm install -g github:na8ve/na8ve-agent, then start it in the folder you want it to work in.
- Sign in. Type /login and choose na8ve. Your browser opens here, you approve, and the agent finishes on its own.
- Try a demo. Type /demo clinical. It seeds a case study into a space of its own and asks its question.
Early access: there are no official releases yet. Node.js 22.19 or newer, and git.
Case studies
five demos, all syntheticEach case study is its own repository: a made-up organisation, its records, and one question that no single fact answers. Clone one, seed it into your memory with na8ve agent, and watch the answer come from memory with its citations.
Thirty conditions and a cohort of patients. One arrives tired, with a high ferritin and bronze skin: which explanation ties the findings together?
A bioreactor, a pipetting arm and a chromatography suite streaming live. When the arm raises an alarm, the memory knows what to check.
A small online shop and its code, kept as facts beside the team's decisions. Ask what depends on a function and the answer comes from memory.
A middle school's course as a map of concepts, and every student's records. Why does one student make three different mistakes, and what comes first?
A binder campaign that stalled, eight instruments streaming a culture, and five plates left. The reason is in the record, in pieces.
Everything in them is generated. No real patient, student, shop, lab or instrument is described.