Context, Semantic Layer, Corporate Memory (What They Actually Mean)
These terms show up in every other LinkedIn post about AI. Everyone uses them slightly differently. So here's how we think about them at Keboola, and what they look like in practice.
Context
Context might be the simplest one. It’s any information that helps an AI agent understand your specific situation.
Without context, you get generic answers. Ask an AI "why did this job fail?" and it'll give you a checklist of common causes. Give it your an error log, your transformation code, and the table schema it was trying to read - now it can tell you the column was renamed last Tuesday and the query still references the old name.
In Keboola, context works on several levels. Kai (our embedded AI agent) automatically knows your project: your screen, your tables, schemas, transformations, flows, job logs, and configurations. You don’t need to copy-paste error messages or explain your data model. Kai reads it directly.
But automatic context only goes so far. Kai knows your schema but not your business rules. It can see a column called margin_adjusted but it doesn’t know why it exists or how your team decided to calculate it. That’s where the next two concepts come in.
Semantic Layer (and Business Glossary)
At its simplest: a semantic layer is the agreement on what your numbers mean. When someone says "revenue" or "active customer," does everyone in the company calculate it the same way? Usually not.
The business glossary is the starting point: a human-readable part. A register of your terms and KPIs with plain language definitions. Most companies keep theirs in an Excel file that someone wrote two years ago and nobody opens anymore.
A proper semantic layer goes beyond that. It's a living catalog where every metric has a definition, a formula, an owner, a threshold - and a connection to the actual SQL that calculates it, so "revenue" IS computed the same way every time, not just described the same way.
Michal Hruška showed what this looks like for finance teams at our webinar.
Another level are thresholds. "Margin should be between 15% and 45% for this product line." With thresholds in place, your agent knows what's normal and what should be flagged, without you having to check every number yourself.
Jiří Maňas, our CTO, uses a good analogy in Episode 1 of our podcast AI-Ready Finance:
You wouldn’t hire someone straight out of university and expect them to know how your company calculates margin. You’d onboard them. Give them context.
AI is exactly the same. The semantic layer is the onboarding.
In practice in Keboola: Custom Instructions are one way to start building your semantic layer. Define at the project level what your terms mean, and Kai follows those definitions in every conversation. Have Kai update your table metadata with business definitions. Every future query uses them automatically. When someone asks Kai "what's our churn rate?" it calculates it the way your glossary defines it, not the way a Wikipedia says.
Corporate Memory
This can be understood as part of semantic layer as well. But it is usually the hardest to solve.
Corporate memory is the knowledge that lives in people’s heads, not in any system. Every organization has someone who knows why the Q3 number looks different from last year. Why that market is excluded from the consolidation. What changed in the data model in 2023 and why. Why the formula in cell D47 of the master spreadsheet has a hardcoded exception.
When that person leaves, the knowledge leaves too. A very specific example story comes from our recent finance breakfast in Prague. A CFO from a manufacturing company described his ideal future: an AI agent that knows why vinyl production costs went up by X% last month. The answer? Summer was coming, the warehouse temperature rose, and the vinyl material behaves differently under higher heat and humidity. Shrinkage rates change, waste increases, production takes longer.
No AI agent would figure that out on its own. It's the kind of knowledge that one production manager knows because they've been working there for 15 years.
Corporate memory is capturing that knowledge: the "why behind the why", and making it available to the systems that need it. Not just humans, but now for AI agents that will increasingly be asked to explain variances, forecast costs, and flag anomalies.
In Keboola, we're rolling out the ability to upload a whole documents directly into Kai's memory inside the platform - think of it like Project Files in your Claude menu, but for your Keboola project. Upload your business notes from a meeting, process documentation, or metric definitions, and Kai draws from them in every conversation.
Jakub Žalio, Group CTO at Creditinfo, talks about this in our podcast AI-Ready Finance Episode 1. Creditinfo operates across 30 markets, each with different regulations, different data formats, different business rules. They spent nine months building their data foundation before turning on any AI.
How They Fit Together
I think of it as layers:
Context is the broadest: everything the AI knows about your situation right now. Your tables, your schemas, your current page your mouse is on in the platform, the error log from this morning.
Semantic layer is the governed meaning of your data. It includes your business glossary (the human-readable definitions of your metrics / KPIs) plus the technical metadata, thresholds, and code that enforce those definitions. The glossary is the content; the semantic layer is the content + the system that keeps it true.
Corporate memory might be the broadest layer: the institutional knowledge that explains why things are the way they are. The business decisions, the exceptions, the history. It's the context that's hardest to capture because it usually unstructured.
Simply stated, an AI agent without context gives you generic answers. With context, it gives you relevant answers. With a semantic layer, it gives you consistent and verified answers. With corporate memory, it gives you answers that actually explain something.
How to Apply it Yourself
You don’t need to build all three layers before AI becomes useful. Start with one. I sorted these by time investment because I know how lazy we all are:
- In 5 minutes: Set one Custom Instructions for your Kai in your project. Your SQL preferences, your naming conventions, your preferred language.
- In about 15 minutes: Configure Tool Permissions so Kai can auto-approve read-only actions. No more coming back to a stalled analysis because you didn’t click “approve.”
- Investing around 30 minutes: Ask Kai to document one of your flows. Then add the business context on top: why this flow exists, what business question it answers, what exceptions it handles.
- Rolling out soon: Upload your glossary/docs into your Keboola Project to give Kai deeper understanding of your company. Go to Project Settings → Kai Assistant → Context files.
Resources
Karolina Everlingova
Product Marketing Manager
