The context plane

Your agency's knowledge is not an asset until it is queryable

Right now it lives in three places: the head of the strategist who is on holiday, a slide deck from March, and a spreadsheet someone renamed. None of them are queryable, none of them carry a date, and none of them can tell an agent what it may safely touch.

That is fine at five clients. At thirty it is the thing capping your margin — because every new hire, every new account, and every AI tool you point at a client starts from zero and rediscovers what you already knew.

A context plane is the layer that holds it: the strategic entities of a business, typed, dated, sourced, and served to any agent that asks. Everyone in this market now claims an agent. Nobody claims a ledger.

This page is the published schema of what ours holds, per client — so you can check your own stack against it.

Do the arithmetic on your own shop

Not ours — yours. Four numbers you already know, and a total nobody has ever put on a slide:

How many hours does a new account manager burn before they are genuinely useful on an inherited account? Multiply by how many people you onboarded last year.
How many audits you already paid for have been redone because nobody could find the first one, or trust it was still true?
What is one renewal worth, where the reason you lost it was that nobody could reconstruct why you recommended what you recommended?
How many AI tools have you bought that you cannot point at a client account, because two of your clients compete?

That total is what the missing layer costs you a year. It appears on no line of your P&L, which is exactly why it never gets fixed — and why it grows with every client you add.

Where it shows up

A resignation letter stops being an invoice: Your best strategist gives notice on a Tuesday. What leaves with them: every objection they heard on a sales call, the positioning they tried in Q1 and quietly dropped, the reason they killed a channel that looks fine in the numbers. Their replacement re-learns it over six weeks, at your cost, and gets some of it wrong. Written down as typed, dated records, none of that walks out. The handover is a read, not an excavation.
You stop paying twice for the same audit: You win an account off another agency. Nobody can tell you which tags fire, which automation depends on a spreadsheet in someone's Drive, or who still has admin on the ad account. So you scope an audit — billable once, filed, forgotten, and redone next year because nobody trusts that it is still true. Held with a stated shelf life, it is done once and tells you itself when it has gone stale.
You can survive the renewal question: Renewals are rarely lost on this month's numbers. They are lost when a client asks why you moved the budget in March and the honest answer is that the person who decided has left and the deck is in a folder nobody can find. When decisions supersede rather than vanish, you can show what you believed, what replaced it, and when — which is a different conversation from “we'll look into it.”
You stop being blamed for numbers you never instrumented: A purchase tag breaks in week two. It produces a figure sitting perfectly between last month and next, so nothing alerts, and the first anyone says about it is your client asking why revenue fell off a cliff. You then spend a fortnight proving it was their developer's deploy. If the signal's trust state was declared — with evidence, an owner and a date — that fortnight is a two-line email.
You can finally point AI at client work: Most agencies have bought AI tooling they cannot actually use on accounts, and the reason is rarely capability. It is that two of your clients compete, and nothing in the stack can guarantee which one a tool was looking at. Every read here names the client it is for and is refused on any doubt; business units carry their own ceilings. That guarantee is the thing that has to exist before any of the rest is usable.
Client ten costs less than client one: Right now every new account starts at zero, so thirty clients cost roughly thirty times one. When a finding shape recurs across engagements it becomes a pattern, and the next client starts from what the last one taught you. That is the only version of agency scale that is not just hiring.

“We have this already. It's in Notion.”

Often true, and a wiki is a real improvement on nothing. It fails three tests, and the third is the one that will matter to you within the year:

Can you query it? Search finds a page. It cannot answer “which of my clients has an unverified competitor claim.”
Does every claim carry a date? A page edited in March and a page still true in March look identical. Nothing goes stale on its own and tells you.
Can an agent read it without a human pasting? This is the one. A wiki is written for people; an agent needs typed records with an isolation boundary. That is the difference between a document and a plane.

We are not claiming a wiki is worthless. We are claiming it stops being enough at the point you want anything automated to touch a client account.

Your reporting tool answers what happened. It holds metrics, dates and dimensions, and it holds them well. What it does not hold is what any of it means: which customer you decided to stop chasing, what the offer promises this quarter, which conversion tag is quietly broken, what you believe about a competitor and when you last actually looked.

The canonical data models in marketing data integration are metric-shaped by design, and the good ones are very good at it: they normalise what platforms measure so a number means the same thing across sources. That is a different job from this one. Everything below is what no metric-by-date-by-dimension store has a shape to carry — and we publish it so the difference is checkable rather than asserted.

The entity map

Each entry: what it is, and how we know it. Provenance is where the fact came from. Freshness is stamped on every entity; a declared budget means the entity goes stale on a clock and says so. Verification is how a claim earns the right to be load-bearing.

Identity & strategy

Entity	What it is
Client	One legal entity, one contract shell — the isolation boundary every read and write is checked against.
Business unit	An operational slice under one client: its own commercial model, team, market or sensitivity ceiling. Most strategy entities are BU-scoped, never client-wide, because “which persona?” usually answers “depends which one.”
Persona / ICP	A marketing persona — segment, psychographic anchor, job-to-be-done, buying stage, and an explicit who-this-is-not-for. Distinct from your contact roster: this is who you’re for, not who’s in the CRM.
Offer	What is being sold and why anyone would buy it — packaging, price, risk reversal, urgency.
Positioning & belief profile	Where the business stands against whom, and the identity, capability and value beliefs of the people who decide its marketing. The layer that explains why a strategy that should work doesn’t.
Messaging house	The message tree — umbrella down through pillars to supporting and objection-answering messages — scoped to a business unit, with locale variants. Every message carries proof-point citations, each one evidence-classed. A message that answers a logged objection points at it.
Objection ledger	Every objection actually heard in a sales conversation, with first-heard and last-heard dates as event time and a source citation on every entry. Closes the loop back into the messaging house.

Epistemics. Provenance is operator elicitation and ingested material — calls, notes, documents — and nothing here is written without explicit operator acceptance. Prose stays canonical throughout this group: some entities also carry a structured record for querying, others are held as the document alone — in both cases the document is the version a human owns and can override. No freshness budget; these change on decision, not on a clock.

Estate & accounts

Entity	What it is
Accounts & domains	The ad, analytics and platform accounts, the domains they serve, their integrations, their dependencies, and who has access to each.
Channels	The marketing channel catalogue — paid, organic, email and the rest — scoped to a business unit and a domain, because “organic” on two domains is two channels, not one.
Measurement stack	The conversion-tracking configuration and conventions: tag-manager and server-side containers, the tag and conversion-action and pixel inventory, and consent state. This is the entity with no date axis at all — it is a configuration, true until someone changes it, and no metric-by-date store has a shape that can hold it. Declared freshness budget: 30 days.
Automation estate	The web of marketing automations, cross-app zaps, the spreadsheets they secretly depend on, the forms that feed them and the tags they set. One record per connected account. Declared freshness budget: 60 days — and an agency handover invalidates it wholesale.
Product catalogue	The synced product feed: items, their attributes, and soft-deletion when an item drops out of a sync rather than vanishing from history.
Reference assets	The brand and campaign asset library indexed by role — logo, product, lifestyle, prior creative, background, locked key visual. Carries a rights gate that defaults to not cleared for ads.
Dashboards	Bindings to the external dashboards a client actually reads, so a number on a slide can be traced to the board it came from.

Epistemics. Provenance is part discovered by direct read of the live account, part operator-supplied — and the record says which. Asset sources are ranked, so a client-supplied logo always beats a scraped one. This is the group that carries freshness budgets, because it is where the world changes underneath you without telling you.

Competition

Entity	What it is
Competitor organisation	A competitor as an organisation, per client.
Competitor edge	The relationship between a competitor and one business unit — never the whole client. You do not compete with a company; a slice of you competes with a slice of them.
Observation ledger	One immutable record per competitor, per layer, per surface, per moment in time.
Candidate queue	Discovered domains that might be competitors, each with a disposition. Discovery proposes; the operator admits. Nothing on the discovery path can create a competitor. Rejected candidates persist with a re-test interval — otherwise the same rejected domain returns as “new” every cycle and the loop never converges.

Epistemics. The observation ledger is append-only: an update is refused outright, and a correction is a new record with a later observation time. Every verdict is three-valued — data, no data, or could-not-read — never a boolean, and a no-data verdict is rejected unless it carries a record of the attempt that produced it. “The competitor doesn’t do this” and “we got blocked” are different facts; a boolean spells them identically.

Performance memory

Entity	What it is
Metric definition	What a metric means and how it is computed, for this client. Read before anything computes, so a metric is defined once rather than guessed anew each time.
Dataset definition	What each connected source actually represents — the “define your sources” layer that every downstream number depends on.
Baseline	A reference point for a metric with a seasonality-aware band, not a single number, tagged with whether it was measured, stated, or provisional, and stamped with the period it describes.
Metric observation	An append-only, event-time fact: a value recorded verbatim from its source, carrying a provenance line back to that source.
KPI	A target for a metric over a window. Its plan value may only come from a campaign, an objective, or an operator — never fabricated, and the field recording where it came from cannot be left empty. Status is computed at read time, never stored, and not enough volume to say is a first-class verdict alongside on-track and off-track.
Objective	The objective layer; key results are KPI records pointed at an objective.
Measurement trust	A per-signal declared trust state — validated, known-gaps, structurally unreliable, unverified, or broken — with evidence, a named owner, and a review date.

Epistemics. Observations are append-only and carry event time, not the time we learned them. Trust state is declared, never inferred — a detector cannot find a broken signal, because a broken purchase tag produces a number sitting perfectly between last month and next. Trust then propagates down the derivation chain by weakest link: if the purchase signal is untrustworthy, so are conversion rate, cost per acquisition and return on ad spend — automatically, without anyone remembering. A stale declaration degrades toward unverified and never upgrades, and a missing declaration refuses the headline figure rather than defaulting to “fine.”

Content

Entity	What it is
Content corpus	The ingested content, chunked, searchable by both keyword and meaning, with every chunk pointing back at the document, meeting or paste it came from.
Content calendar	Planned and published pieces, tied to the persona each is for.
Topic cluster	A cluster of intent scoped to a market cell — business unit, language, country. One target page per cluster per cell, enforced: a second cluster claiming the same page in the same market is rejected. Where a cluster states an aggregate search volume, it must also state its inclusion rule and its exclusions — a total whose exclusions are unstated is two teams quoting different numbers for the same cluster.
Keyword	Unique per term, language, country and business unit. The same word in two markets is two keywords, because it is two different questions.
Rank observation	An append-only rank fact: one record per keyword, device, engine and date.
Page snapshot	A full page as it was at a moment, so an analysis can be re-run against what was actually there.

Epistemics. Market scoping is structural, not a convention people are asked to remember. Rank and snapshot records are append-only and keyed by observation date, so a re-crawl never overwrites history.

Decisions & governance

Entity	What it is
Finding	A diagnostic observation — something true about the business that someone established.
Recommendation	What to do about a finding. It points at the finding it addresses, and the finding points back.
Pattern	A finding shape seen across more than one engagement — the record that turns one client’s lesson into the next client’s head start.
Engagement memory	The working memory of the engagement: decisions, notes, briefs, transcripts and completed outputs, searchable by keyword and meaning.
Knowledge graph	Typed edges between all of the above, each carrying a confidence and a status — so a machine-proposed link and a human-confirmed one are never read as the same strength of claim.

Epistemics. Findings supersede, never delete — a corrected fact points at the record that replaced it, and the replaced record stays readable. Engagement memory is bi-temporal: when a fact became true is recorded separately from when we learned it. Every memory carries a verification state — unverified, verified, refuted — and only verified memories are load-bearing; causal memories carry their ruled-out hypotheses beside the surviving one, so retrieval can surface why, not just what. And every memory is provenance-classed by whether it came from the world or from our own earlier output — because two sources both downstream of something we wrote are not two sources, and a system that cannot tell will confirm itself into confidence.

What makes a fact here different

Evidence classes: Every material claim is tagged with how it is sourced: a number from a system, something we verified ourselves, something the client stated, a repeated story, an inference, or hearsay. A seventh marker exists for a read that was attempted and failed — a fact about the instrument, not the subject. Only classes that can be independently checked may carry a conclusion; a story told five times does not become data.
Provenance to the source: A number carries the specific account it was read from, not the vendor's name. Several accounts answer to the same vendor, so a stamp reading just the vendor looks like provenance while failing to be it. Every fact resolves to a source you can open.
Freshness, stamped and budgeted: Every fact records when the event happened, separately from when we fetched it. The event time is computed from the payload's own timestamps with no path to today's date, so “the request succeeded” can never masquerade as “the data is current.” Where a budget is declared, the fact reports fresh or stale-by-N-days; where none is, it reports unknown — never fresh.
Three-valued verdicts: Every read resolves to one of three outcomes: we got data, there is genuinely none, or we could not read. A zero from a successful read is evidence that something is dark and needs fixing; a zero from a failed read is evidence of nothing. Collapsing those into a boolean is how reports lie quietly, every month.
Correction, not deletion: The observation ledgers refuse updates outright — a correction is a new record with a later timestamp. Findings supersede rather than disappear. You can always ask what we believed in March and why we stopped.
Isolation at every call: Every request through the platform's interfaces names the client it is for and is checked against that client's access grant before anything is read — and is refused on any doubt. Business units carry their own sensitivity ceilings, so someone granted one slice never sees strategy they were not granted.

How you reach it

The context plane is headless and served over MCP, so any agent surface can read it — your own assistant, your team's copilots, ours.

Arkana, the in-built agent, answers from it directly: it routes a question deterministically, answers from memory and the structured record where it can, and only then spawns a bounded run. Every answer carries its citations, its freshness codes, and the sets it searched. A question it cannot ground returns a typed gap naming what was missing — never a guess that reads like an answer.

The check we'd like you to run

Pick your largest account — the one you would least like to lose — and ask your current stack three questions about it:

What is the consent state of that client's tracking configuration right now?
Which competitor claim are we holding for them that nobody has re-verified this quarter?
Which of last quarter's decisions was superseded, and what replaced it?

If it cannot answer — and it likely cannot, because none of those are metric-shaped — then the answers live in someone's head, and you are one resignation away from losing them. That gap is what this layer is for.

Design partners

Turn what your team knows into something the next hire can query.

We're taking on a small number of design partners: agencies, fractional CMOs and in-house marketing leads who want their strategic context to be an asset on the balance sheet rather than institutional memory walking out of the building.