Mitosis Cortex
Your agents stop guessing.
Cortex is the source of truth your agents read from. It turns the email, chat, documents and records your company already has into one structured memory, so an answer comes back from what is true about your business rather than from what sounds plausible.
Free to start. A paid Cortex plan includes Yappy.
A colony you can poke at
This is a real Cortex colony, drawn from the live database rather than a screenshot. Every dot is something the extractor pulled out of a source, every line a relationship it worked out. Drag it, zoom it, hover a node to light up what it is connected to. It is the public Grokipedia audit graph, which is the one colony we can show you without showing you somebody else’s company.
Your own colony is not this one. Every workspace runs in its own namespace with its own database, and there is no query path from one to another. This graph is public because we made it public.
Measured, not asserted
These are the published figures. Both runs are reproducible through the public API, and the audit was run by someone who does not work here.
Independent evaluation: Verging Labs, Cortex on the Agentic Memory Index v0.1, window 23 to 25 July 2026. The store was built and probed entirely through the public APIs, using a deterministic matcher that escalates to an LLM judge calibrated against human labellers. Full method and the LongMemEval run are on the research page.
What it actually is
Think of the filing system a good assistant leaves behind when they move on. The work still gets done, and it gets done the same way, because the structure survived the person. That is the job Cortex does for software.
Most AI disappointment traces back to the same place. The model is fine. It just has no memory of your company, so every session starts from nothing, a person re-explains the context, and the output still has to be checked. The checking is the cost. Teams end up paying for the AI and then paying themselves to audit it.
Cortex removes the re-explaining. Consistency stops depending on which model you picked this quarter.
How it works
Three steps, and the third one is the one that pays for the other two.
Connect your agents
Claude, ChatGPT, Gemini, Hermes, OpenClaw or anything that speaks MCP. Cortex becomes the memory they read from. Nothing about how your team already works has to change.
Connect your sources
Email, chat, documents, drive, Slack, WhatsApp, GitHub, Salesforce. Each connector reads and writes, so an agent can file something back rather than only looking things up.
Stop paying to re-derive
Once an answer is indexed, the next agent that needs it looks it up instead of reasoning its way there again. That is where the spend goes down and the consistency goes up.
What buyers ask
The six questions that come up on every evaluation call, answered here so you do not have to book one to hear them.
Accuracy
Independently evaluated. 0 fabricated answers and 0 flat wrong answers across 272 scored questions. On facts it was never given, it says it does not know 97 percent of the time, which is the behaviour that makes a team trust it twice.
Isolation
Every workspace runs in its own namespace with its own database and its own pods. Your graph is not a row in a shared table, and there is no cross tenant query path to get wrong.
Provenance
100 percent of answers carry a source receipt. You can click through to the message, document or record an answer came from, which is the difference between a system a compliance team can sign off and one it cannot.
Cost of ownership
The expensive part of agent work is asking a frontier model to rediscover what your company already knows. Cortex moves that lookup off inference, so a smaller model does the same job. Teams keep the model they like and spend less running it.
Adoption
The blocker on internal AI is rarely capability, it is trust. People stop using an assistant that has confidently wrong days. Refusing to guess is what keeps them coming back, and it is why retention is the metric we watch rather than seats.
Compliance
SOC 2 audit is funded and in progress, not yet complete. Snowflake migration is underway. We would rather tell you that now than at procurement.
Multiplayer
One memory, many people writing to it. When your colleague files something, your agent knows it, without anyone forwarding anything. Wikipedia asked for this before we built it, which is the best argument for a feature we know of.
Shared memory only works if new information cannot quietly overwrite established fact. Cortex weighs how recent a claim is against how well supported it is, so the loudest recent input does not win by default. Redundant and weakly connected data is pruned on a schedule, which is what keeps both the accuracy and the bill from drifting upward over time.
Pricing
Usage based, so the bill tracks what you actually ran.
Start
Free
no card to connect an agent
Connect an agent and a source, and see whether it answers better. Most of the value shows up in the first few days.
Team
Usage based
includes Yappy
Shared multiplayer memory, every connector, source receipts on every answer. A paid Cortex plan includes Yappy at no extra cost.
Enterprise pilot
$7,500
fixed scope, fixed price
We map your systems, set the privacy boundaries and report what your current AI spend is actually buying. Sized to come out of the AI budget you already have.
Point one agent at it.
You will know inside a week. That is the whole evaluation.