The company brain
A brain your agents can actually read.
Our company brain is a folder of plain markdown. It holds what no other system can: who people are, what was said, what was decided and why. People write in it. Agents read it, add to it and keep it tidy.
Ask it something
Watch a question travel through the brain.
Every dot is a note. Every line is a link. Pick a question a consultancy gets asked every week and follow the signal: which notes fire, in what order, and the answer the agent writes from them.
- Records
- Compiled notes
- Generated
Route
- 01Read the use case indexIndex/Use Cases Index.md · one line per use case
- 02Four descriptions matchno reads: the index line is the description
- 03Open the closest match, follow its linkKnowledge/Use Cases/Arc Health: Looker to Omni…
Answer
Yes, four times. At Arc Health we guided a 30-person data team through moving about 500 business users, using audits and scripts to decide what to keep, then training in waves. Lucky Quay was several hundred dashboards, with their engineers and analysts trained to run it themselves. Relayo took two weeks of daily co-development and went live before the old tool's deadline. Sendwell was customer-facing embedded analytics under a contract deadline. We also built a migration tool at Arc Health and have reused it since.
Route
- 01Read one generated decision logDecisions/By Company/Claimbridge: Meeting Decisions 2026.md
- 02Each decision points back to its meetingno reads: the links are the evidence
Answer
Latest first. On 16 Oct we agreed to keep Corvane UAT live alongside Prod, and that the KPI definitions are final. On 16 Sep we cancelled the SLA report ticket, because the SLAs cannot be reported reliably from Monday.com today. On 7 Sep we agreed UAT runs at Claimbridge's pace, with no pressure applied. All of it comes from one 5KB log that a script builds from the meeting notes, so nobody had to open eleven of them.
Route
- 01Find the company in the indexIndex/Companies Index.md
- 02Open the company noteCompanies/Norhaven Logistics.md
- 03Read its People blockgenerated from 69 meetings, with a count per person
Answer
Five people to know. Hamish Brackenbury is the COO, with 13 meetings on record. Gordon Harding is the CEO. Edmund Varley leads financial modelling and the app build and is our most frequent contact, at 32 meetings. Elijah Okafor is in finance, and Avery Hallett runs portfolio management. The counts come from links in the meeting notes, so they are current as of the last build.
Route
- 01It is a tool question, so open the tech noteTech/Omni Embed.md
- 02Follow its fixes sectionthree linked notes, one gotcha each
- 03Check where it was learnedReferenced by: 11 use cases, 2 meetings
Answer
Three, all learned on real builds and collected in one tech note. Tiles that are not backed by a Topic fail for restricted queriers, and the fix is accessBoost. There are three ways to give an embedded user a dashboard (signed URL, shared folder, entity folder) and each suits a different case. Template dashboards can be seeded into every entity folder through the duplicate API. Most of this comes from the Verico embed build, which runs a persistent iframe session with entity folders and connection roles.
Route
- 01Read the properties onlypipeline_company_id: 2d802f18-…
- 02Hand off to the system that owns it$ pipeline opportunities get <id>
Answer
The brain does not know, and that is deliberate. It holds the id of Norhaven's pipeline record and hands the question to the pipeline, which answers with today's stage and value. A deal value copied into a note would be wrong by next week.
- 01Index
- 02Note
- 03Links
- 04Search
- 05Transcript
Cheap first. Stop the moment you have the answer. Most questions end on rung two or three.
Demo data from an anonymised copy of our own vault. Every client, company and person named is fictional, and figures are altered.
- notes, one type each
- 1,050
- links between them
- 16,619
- decisions surfaced from meetings
- 729
- lint errors allowed
- 0
Structure
Three layers. The layer decides who may change a note.
Every note sits in exactly one layer. That one rule stops the two things that rot a knowledge base: history being rewritten, and summaries going stale.
Records
Dated, and not rewritten after the fact. A meeting record from June stays the June record.
- Meetings/
- Sources/
- Journal/
- Decisions/
If it turns out to be wrong, add a dated correction or write a newer record that supersedes it.
Example: Meetings/2026-06-01 Norhaven AI Sprint Plan.md
Compiled notes
Present tense. Rewritten freely when the evidence changes. One note per person, company, project, tool or method.
- People/
- Companies/
- Projects/
- Tech/
- Knowledge/
Every claim traces to a record by link: "Edmund owns the Python asset models (since Jun 2026, [[meeting]])".
Example: People/Hamish Brackenbury.md
Generated content
Timelines, company maps, decision logs and indexes. Nobody edits them. A script rebuilds them from links and properties.
- Index/
- Decisions/By Company/
- Home.md
- generated blocks
If code can derive it, code maintains it. No hand-written meeting lists, backlinks or indexes.
Example: Index/People Index.md: every person, one line each
Anatomy
Every note says what it is before you open it.
The properties at the top are the retrieval surface. Pick a line to see what it buys the agent.
type
One value, and it must match the folder. The path already told the agent what this is before it opened the file.
description
The line an index shows. An agent reads this, not the note, to decide whether the note is worth opening. Written as the answer to "why would I open this?".
relationship
Short, reused property names mean one filter works across every company: customer, prospect, partner, supplier.
System ids
Written only by the sync script, which checks each id against the live database first. This is how the brain points at live data without copying it.
Derived fields
Computed by the build from meeting links: date of the latest meeting, and the meeting count. Nobody types these.
aliases
Every nickname, former name and known mis-transcription. Search finds the note whatever the call recorder heard.
Links
The first mention of any person, company, project or tool is a link. Links are the graph: timelines, maps and "referenced by" lists are all computed from them.
Generated block
Everything between these markers is rebuilt by the script on every run: the 30 latest meetings, people, projects and decision logs for this company. Hand edits here are overwritten.
Generated block
Everything between these markers is rebuilt by the script on every run: the 30 latest meetings, people, projects and decision logs for this company. Hand edits here are overwritten.
One line per record
Each line carries the meeting description, so the agent reads the outcome of 69 meetings without opening one.
Writing
The model judges. The script writes.
After a meeting, an agent runs the same seven steps every time. The model does the parts that need judgement. Code does everything that can be derived, so no one hand-keeps a list.
Write the meeting record. One note in Meetings/, every property filled. The description is the outcome in one or two sentences, not the agenda.
File the transcript. The verbatim transcript goes to Sources/Transcripts/ and is linked from the record. It is immutable, and out of the default search path.
State the decisions. Each decision is one bullet, written to stand alone: it names the thing decided, never "agreed to do it". These bullets become the per-company decision logs.
Create person notes. Every attendee without a note gets one, with their company set. A role is recorded only if a source states it. It is never inferred.
Update what is now true. Only if the meeting changed something. Rewrite the relevant paragraph of the person, company, project or tool note, link the meeting as evidence, and set updated.
Sync, build, lint. vault sync-ids links new notes to the pipeline and ledger. vault build regenerates timelines, maps, decision logs and indexes. vault lint checks every note against the schema. Errors must be fixed before finishing.
Log it. One line appended to Log.md: date, skill, what was written. The log is append-only, so there is always a trail of who changed the brain and when.
Boundaries
The brain never holds live data. It holds the id.
Deal values, tasks, hours and balances change daily, so they stay in the systems that own them. A note carries the id of its record, which turns “look it up” into one exact query and never a name search.
Pipeline
Deals, stages, values, contact details
vault keeps: pipeline_company_id
Linear
Tasks and tickets
vault keeps: a project URL. No checkboxes.
Project ledger
Time and billing
vault keeps: ledger_project_ids
QuickBooks
Money
vault keeps: why, never how much
Google Drive
Shared documents and deliverables
vault keeps: a link
GitHub
Code and technical docs
vault keeps: a link
# "What stage is the Norhaven deal at?" is not a vault question.
$ pipeline opportunities get 2d802f18-db20-8781-3093-9c7451c30807
$ ledger projects status "Norhaven Web App Support"
- The description is the most important field
- One or two sentences answering "why would I open this?". It is what an index shows and what an agent reads to decide. Capped at 300 characters.
- Links, never tags, for things
- A client is
companies: "[[Norhaven Logistics]]", not a tag with three spellings. Links are what the generated blocks are computed from. - One name, one note
- Names are unique across the vault. People are linked by full name. An unresolved link is fine: it marks something worth a note later.
- Size limits keep reads cheap
- Compiled notes stay under 25KB. Generated lists show the latest 30 entries. Transcripts live apart and are opened last.
Want a brain your agents can read?
We turn how your business thinks and operates into a brain AI can use. Tell us about your business and we'll map where to start.

