How It Works
The Company Brain has two parts: knowledge of what your data means, and memory of what your agents learn. Data feeds both, and each answer can draw on both.
The Company Brain has two parts that work as one: knowledge (what your data means) and memory (what your agents experience). Everything you bring in feeds both, and every question can draw on both.
The end-to-end flow#
Two streams feed the brain. Your data sources flow in through connectors, and your agents contribute as they work, recording what they learn and are told. Both enter through the same path.
Everything you bring in is processed in the background and fans out across the brain at once: it becomes searchable shared memory. When that content carries structure, it also enriches the semantic knowledge layer. You ingest once; the brain puts it to work everywhere.
The brain doesn't just store. It distills memory into events and the relationships between them, and maps your data into the Ontology Graph, weaving both into a connected understanding of what your company knows and has done.
Any agent asks a question in plain language. One query reaches across both knowledge and memory at once, runs whatever steps are needed, and returns a grounded answer with citations and a confidence signal, within the access that agent has been granted.
Corrections and completed work flow back into the brain, sharpening both its knowledge and its memory. Because the brain is shared, every improvement reaches every agent. The more your company uses it, the more useful the brain gets.
One brain: knowledge and memory#
The brain's power is that these two halves are not separate systems bolted together. They are one system. A single question can weave a fact from your data together with something an agent learned last week.
What your data means. The Ontology Graph turns your connected sources into a semantic map, so plain-English questions reach the right data and come back grounded.
What your agents experience. Event-driven shared memory keeps what your agents learn, decide, and are told, available to the whole company and never trapped in a single session.
Three kinds of memory#
When the brain answers, it draws on three distinct kinds of memory:
What is objectively true about your data: the structure of your connected sources and documents and the relationships within them. Read directly from your sources, and re-synced when your data changes. It cannot be "wrong" the way an inference can.
What the brain has learned: business concepts inferred from your data, and know-how distilled from completed tasks. Persistent and continuously evolving, refined over time as the brain verifies it.
Agents can remember past conversations and what previous agents did. Working memory holds the current chat, instructions, reasoning, tool outputs, and partial results while an agent works; useful parts are saved back into persistent memory.
Experiential memory is where self-learning happens. Chats, completed tasks, and agent outputs are summarized and saved so future agents can reuse them. See Memory & Knowledge Layer.
A worked example#
An agent asks: "What was last quarter's revenue by region, and how does it compare to the target we set?"
- From knowledge, the brain finds "revenue" mapped to your data and aggregates it by region.
- From memory, it recalls the target the team agreed on in a planning conversation last month.
- It runs the comparison, holding the intermediate work in working memory.
- It returns one answer that uses both: the live numbers and the remembered goal, with the sources it drew on and a confidence signal.
- If someone revises the target later, that update flows back into memory and improves every future answer.
This is the integrated brain at work: the agent never had to know where "revenue" lived or remember the target itself. The brain held both and reasoned across them as one.
The Data API streams each of these stages back as it works, so you can show reasoning progress in real time.