Watch a thought travel through the brain.
Ask it something and the whole mechanism runs in the open: what it takes the question to be about, which moments it pulls back and why those, the part of your company it lights up, the checks it runs, and the answer with its receipts. Nothing below is a metaphor. It is the loop, slowed down enough to read.
Pick a question and watch the turn build: what it reads, what it remembers and why, what it lights up, what it checks, and the answer it will stand behind.
One question. Five phases.
Nothing happening off-screen.
The demo above runs the same five phases the product runs. Here is the anatomy: what each phase takes in, what the brain actually does with it, and what it hands to the next one. It is a single path a question travels, not five features stacked in a box.
reads the question and picks out what it is about
Your question, in the words you actually used.
Breaks it into the things it is really about: the metric, the region, the cost driver, the time window.
A short list of entities, held in working memory for the rest of the turn.
pulls back the moments that actually bear on it
Those entities, plus wherever this conversation had already got to.
Scores the episodic log on meaning, recency, importance and how strongly the moment was felt, then takes the top of that ranking.
The handful of moments that bear on this one, each carrying the four numbers that surfaced it.
lights the part of your company the question touches
The entities and the moments that came back.
Lights the part of your company graph the question touches and walks the drivers backwards, one edge at a time.
The nodes, edges, owners and agreed definitions the answer is obliged to respect.
checks the data, and looks things up when it must
A claim worth testing, and the gaps it cannot fill from memory.
Queries your own data. When the answer is not in there, it searches the public web, opens the page and reads it.
Figures, passages and a screenshot of what it opened, all kept as evidence.
answers, and shows you what the answer rests on
Everything above, weighed against what it already believes.
Commits to an answer, says how confident it is, and attaches what the answer rests on instead of asking for trust.
A grounded answer with its receipts, and one more episode appended to the log.
Every phase publishes what it produced, so a turn can be read end to end afterwards. If a claim has no path back through these five, it does not get made.
Five kinds of memory,
because one kind answers nothing.
RECALL does not search one pile. It reaches into five stores that answer five different questions: what is live right now, what happened, what things mean, what has worked before, and what the brain currently believes. This is the anatomy underneath the phase you watched above.
What is live right now
A deliberately small hot set: the question on the table, the entities it decomposed into, the goal it is working towards and the thread it is in. It is bounded on purpose. A brain that holds everything at once prioritises nothing, and what falls out of the hot set is not lost, it is in the log.
What actually happened
An append-only, hash-chained log of every signal noticed, every line of reasoning, every conversation, every decision you made, every action taken and every outcome that followed. Each entry carries the hash of the one before it, so the order is provable and a quiet rewrite is not possible. Nothing is edited, nothing is deleted, nothing expires.
What things mean
One graph of your company: sources, tables, fields, metrics, documents, projects, reports, regions and people, plus the edges between them and your own definitions of margin, churn and active customer. This is the store that lets it answer why did this move rather than only what moved.
What has worked before
Playbooks distilled from fixes that verifiably worked. A recommendation whose predicted impact was confirmed afterwards gets promoted into a procedure, so the next time the same shape of problem appears, the proven play is offered first and the reasoning behind it is still readable.
What it currently holds true
Convictions with a number attached: freight is the driver of the West margin move, held at 0.74. Beliefs are not facts and are never presented as facts. They move when outcomes land, up when a prediction is verified and down when it is not, which is what stops the brain sounding equally certain about everything.
Recency alone would hand you this morning. Importance alone would hand you the same crisis forever. Blending all four is what makes the right moment come back.
Hold a node. See what it is attached to.
KNOW is not a lookup. Every source, table, metric, document, region and person lands in one graph, and the edges are the whole point: they are how the brain answers why did this move. Hover a node, or tab into the graph and walk it with the arrow keys. Then run the driver trace and watch it go backwards.
No node selected. Use tab to enter the graph, then the arrow keys to move between nodes.
The edges are the answer
Why did this move is a walk, not a lookup: gross margin ← freight cost ← carrier mix ← West lanes. The brain follows the chain and reports the link that actually moved, with the query behind each step.
Ranked by importance
Nodes are scored, so it leads with the metric that carries the business instead of whichever table happened to load first. The ranking is visible, and you can argue with it.
The gaps are published
Coverage is measured, and what has not been mapped is listed out loud: unmapped tables, terms with no agreed definition, metrics with no owner. A brain that hides its gaps is guessing.
Six numbers that decide
what it opens first.
Concern, confidence, curiosity, satisfaction, stress and urgency, each computed by appraising real events in your data against your goals and what the brain currently believes. They are a prioritization readout, published so you can audit them, and every one of them drills down to the cited events that moved it.
This is not consciousness, sentience or emotion in any human sense. These are six scalars computed from events you can inspect, published so you can check them and argue with them. They exist because a brain with a thousand things it could look at has to choose one, and we would rather that choice was a number on a screen than a hidden ranking.
Borrowed from how thinking is described.
Implemented as ordinary software.
This is an engineering analogy, not a claim to be a mind. Every row on the right is a score, a log, a graph or a threshold that you can open and inspect. Intelain is not conscious, it is not sentient, and it does not feel anything. The parallel is useful because it explains why the parts are shaped the way they are, and it stops being useful the moment it is taken literally.
You do not scan a room evenly. Something pulls your eye first, and everything else waits.
Vitals plus importance scores on the graph decide what it opens first. The ranking is published, so you can see what it put at the top and why.
You can hold only a few things in your head at once, which is why a long list becomes a written list.
A small hot set: the question, the entities it decomposed into, the goal and the thread. Bounded on purpose, and what drops out is in the log, not gone.
You remember the quarter that went wrong far better than the one that went fine. The moments that hit hardest are the ones that come back.
Recall scores how strongly a moment registered alongside meaning, recency and importance, so the March crisis outranks yesterday. Every recalled moment shows all four numbers.
What you know about your company without looking anything up, and how it all hangs together.
One graph of every source, table, metric, document, project, region and person, plus your own definitions, with the edges that make why did this move answerable.
The thing you can now do without thinking, because you did it enough times and it worked.
Playbooks distilled from fixes whose impact was verified afterwards. The same shape of problem gets the proven play first, with the reasoning still readable.
The unease that makes you check the number one more time before you sign.
Six vitals appraised from real events, steering what it opens, when it interrupts and whether it may act. Not a feeling: a threshold with receipts attached.
Overnight, the day gets filed: what mattered is kept, the rest thins out, and the morning feels clearer.
A nightly pass consolidates the day into the log, recalibrates confidence against outcomes that were verified, promotes what worked, and has the briefing built before anyone opens a laptop.
Knowing that you do not know: the itch that sends you to check before you answer.
Measured coverage, a published list of its own gaps, confidence stated out loud, and a read-only investigation loop for the things it cannot answer from what it holds.
The left column is why the architecture looks like this. The right column is all we are claiming.
It investigates.
You decide.
The interesting part of a brain like this is not what it can do. It is where it has been built to stop. There is a line through the middle of the product, and it is drawn in the architecture, not in a setting somebody can forget to switch on.
Investigation cannot write
The research loop queries your data, searches the public web, opens pages and reads them. By construction it holds no path back into your systems, so nothing it discovers can change anything on its own.
A named human approves anything that acts
Every action is risk-classed. Anything that touches the real world is prepared end to end and then stops at an approval that a named, authorized person has to give. High risk needs a human at every autonomy level, with no override.
Both sides are recorded, permanently
Every query, recommendation, approval and action is appended to a trail that cannot be edited or deleted, and real-world actions ship with a rollback plan prepared before they run.
One brain per company
Your memories, your graph, your playbooks, your beliefs and your vitals are yours alone. Isolation is enforced at the boundary and covered by tests: one company’s brain cannot read another’s.
It can think for as long as you like, and it can show its whole working. The decision was always going to be yours.
See it run on your own data.
A 30 minute live session on a seeded company that looks like yours, with the grounding drawer open the whole time.