“We want an AI that knows our business.” It is the most common request I hear, and it hides three different needs. Knowing how many orders shipped last week is one. Knowing what the warranty says is another. Knowing that this client always wants a call before an invoice goes out is a third.
Each has its own tool. Using the wrong one is how AI projects end up confidently wrong.
The database: exact facts
Customers, orders, prices, stock, appointments — records with fields. A database answers “how many”, “which ones” and “how much” exactly, every time, for a fraction of a cent. It is also the only one of the three that can be trusted to change things: record a payment, move a date.
Its weakness is fuzziness. Ask it “something to furnish a small apartment” and it finds nothing, because no record contains those words. An AI assistant connected to it — through MCP, with your access rights — gets the exactness without the fuzziness problem: it translates your question into the precise query.
RAG: what is written in your documents
Manuals, policies, FAQs, contracts, guides — prose. RAG cuts them into passages, finds the few that match a question by meaning, and has the AI answer from those only, citing its source. It shines when the answer is written somewhere and the question can be phrased a hundred ways.
Its weakness is anything that needs every record at once. “What is our total revenue?” or “list all clients in Laval” cannot be answered from five passages. And numbers buried in dense tables are where it slips: in our own tests, the serious errors were values read from the wrong row of a dosing table.
AI memory: how you work
The third kind of knowledge is never written in a manual: decisions, preferences, lessons learned. “Quotes over $5,000 need my approval before they go out.” “Our biggest client pays on the 15th — no reminder before the 20th.” “Never offer a discount on installation.” “Answer Quebec customers in French first.” An AI assistant that keeps short notes like these — an AI memory, sometimes called a second brain — and reads them back at the start of every session, stops repeating the same mistakes and stops asking the same questions.
That is how I work with AI every day: a dozen notes, one fact each, with the reason behind it. Its weakness is size and age. AI memory is read in full every time, so it has to stay small, and a stale note is confidently wrong — so notes are dated, and anything they name is checked before acting on it.
Side by side
| Database | RAG | AI memory | |
|---|---|---|---|
| Holds | Records: customers, orders, prices | Documents: manuals, policies, FAQs | Notes: decisions, preferences, lessons |
| Typical size | Millions of rows | Thousands of pages | A few dozen notes |
| Best question | “How many? Which exactly?” | “What does our policy say about…?” | “How do we do things here?” |
| Answer | Exact | Quoted, with its source | Context for judgement |
| Fails at | Fuzzy questions | Totals, full lists, dense tables | Growing large, going stale |
| Kept up to date by | Your application | Uploading new documents | The assistant, with you reviewing |
The three-question test
- Must the answer be exact — a number, a list, a yes or no? Database.
- Is it written somewhere in prose, and could the question be asked a hundred ways? RAG.
- Is it how you work — something nobody wrote down but everyone should know? AI memory.
Three common mistakes
- A price list in RAG. Prices change and must be exact. Keep them in the database and let the documents explain the policy around them.
- Everything in AI memory. It is read in full every session; a bloated memory is noise, and old notes contradict new facts. One fact per note, dated, pruned.
- Keyword search on a database for fuzzy questions. Visitors do not know your vocabulary. Put search by meaning in front of the records, and keep the records as the truth.
How they fit together
The useful setups use all three, each for its own job. An assistant starts from its memory — it knows your biggest client pays on the 15th, so that client gets no reminder before the 20th. It asks the database, through MCP, for the exact invoices overdue this month. It searches the documents, through RAG, for the late-payment policy. Then it drafts the reminders and waits for your go-ahead.
None of the three is new. What is new is that an AI can use all of them in one conversation — which makes choosing the right home for each piece of knowledge the decision that matters.