A company-wide memory
A shared memory of business logic, requirements, modeling conventions, and expectations that every agent reads and writes. Institutional knowledge as living context.
Institutional knowledge, written down once and shared. Every agent gets smarter as the memory grows.
For most of BI’s history, the knowledge that made the numbers trustworthy lived in people’s heads. Why margin excludes intercompany sales, which customer hierarchy the board actually uses, why last year’s model was built the way it was: a few heads and a scatter of wikis nobody read. When someone left, it left with them.
A company memory is a living record the whole organization writes to and reads from: a queryable store of business logic, conventions, and prior decisions that every agent draws on before it acts and adds to when it learns something. Institutional knowledge stops being tribal and becomes context the system can use.
Why now
Agents are only as good as their context. A capable model with no memory of your business will rebuild the same model three ways and reinvent decisions you settled years ago. The bottleneck is no longer reasoning; it’s grounding.
The semantic layer holds the what: the agreed definition of revenue, the canonical regions, the entities that matter. But definitions don’t explain why: the decision to treat returns a certain way, the convention that every model exposes a date dimension, the expectation that finance figures reconcile to the close. A company memory holds that layer, and it’s what turns a fluent agent into one that fits your organization rather than fighting it. It’s well-trodden ground: capturing decisions and rationale so a project doesn’t relearn itself is the same discipline that runs through a serious migration knowledge base, except now the readers and writers are increasingly agents.
What it looks like
An agent is asked to build a new model for subscription revenue.
Before writing anything, it consults the memory. It finds the team’s modeling conventions (naming, how slowly changing dimensions are handled, the standard date table, the rule that every fact exposes a clear grain) and prior decisions, like recognizing subscription revenue monthly rather than at signing, and why. It builds accordingly, so the new model behaves like the others rather than a stranger’s work.
Along the way it makes a fresh choice, how to handle mid-cycle plan changes, and records it back with the reasoning. The next agent, or the next analyst, inherits that decision instead of stumbling into it. The organization got slightly smarter, and the knowledge didn’t evaporate when the task closed.
Where it’s heading
Toward an organization that compounds its own context instead of paying to relearn it every quarter. As more work runs through agents that both read and write the memory, the record gets richer with every model built and question answered. Onboarding a new analyst, or a new agent, stops being a months-long apprenticeship in unwritten rules and becomes a matter of reading what the system already knows. The asset that used to walk out the door now stays and grows.
How we think about it
Context engineering is the real moat. Models are largely a commodity and improve on their own; what no competitor can copy is the accumulated, curated record of how your business actually works. Memory turns one-off answers into institutional knowledge: an answer with no memory behind it is a clever guess that happens to be right; one grounded in your context is the organization speaking with one voice. So we govern the memory with the same care as the semantic layer and the same discipline as trust and governance: curated, attributed, and trusted, because everything downstream now reasons from it.
A company-wide memory, in short.
How is this different from a wiki or a data catalog?
A wiki is written for people and read by almost no one; it goes stale the moment the author moves on. A company memory is written and read by the agents doing the work, so it stays current because it's used, not archived. The catalog says what a table is; the memory says why it's shaped that way.
Is this not the same thing as the semantic layer?
They're complementary. The semantic layer holds the what, the agreed definitions of metrics and entities. The memory holds the why, the decisions and conventions behind them. One makes answers correct; the other makes them consistent with how your organization actually works.
Who decides what goes in the memory?
The people who already hold the knowledge, except now they record it once instead of explaining it repeatedly. Agents propose entries from the work they do; humans curate. The point is to capture decisions as they're made, not run a documentation project after the fact.
Keep exploring
The semantic layer
The shared definition of metrics and entities, and the single biggest lever for AI accuracy. Answers come from governed definitions, not raw-schema guesswork.
Trust and governance
Accuracy, lineage, access control, and a graceful "I do not know". The precondition that makes every other capability believable.
Where could this take your BI?
If this is the direction you want to head, we should talk.