The demo everyone builds first is a search box over the company drive. Ask a question, get an answer, watch the room nod. It is a good demo. It is not a coworker, and the gap between the two is the interesting part.
#Two things that get called memory
- Retrieval
Finding the document that answers a question. It is a solved shape: index the files, embed the chunks, rank the results.
- Memory
Knowing that Northwind always pays thirty days late, that Marco signs anything over five thousand, and that the discount you gave Halcyon in March was a one-off nobody should repeat.
Retrieval finds a page. Memory carries a rule. Nobody wrote the Northwind rule down in a document called Northwind Rules. Somebody said it in a thread, once, eleven months ago, and everyone who matters has known it since.
#Scope is the feature
The instinct is to give the AI everything, because more context sounds like more intelligence. In a company it is the opposite. A finance coworker that can read the recruitment folder does not become better at chasing invoices. It becomes a question your general counsel has to answer.
So a coworker in Alfera reads the sources attached to it and nothing else. Not "everything, filtered at query time" — the difference is the whole design:
What most products do
Everything, filtered on the way out
The coworker can reach the whole drive. A rule decides what to show for this question, for this person, this time.
The filter is a permission you have to trust the query to apply. Every new kind of question is a new chance to apply it wrongly.
What we do
Only what is attached
The coworker can reach the ledger and the contracts. The recruitment folder is not on the list.
There is no query that reaches it, because there is nothing to filter. It is not hidden from the coworker. It is not there.
That has a consequence worth stating plainly, because it is the part people check: change the scope and it forgets. Remove the ledger from a coworker and the facts it drew from the ledger stop being available in the same second. It does not keep a summary of what it used to know.
A summary of a document you are no longer allowed to read is a copy of a document you are no longer allowed to read.
#Where the rules come from
The Northwind rule is not in a file, so it cannot be retrieved. It gets learned the way a new starter learns it — somebody corrects them.
Don't chase Northwind, they always pay late and they always pay.
That sentence, in a thread, is worth more than the whole payments folder. So we treat a correction as the highest-value thing that happens in a workspace.
Two reasons the citation matters more than it seems:
- You can check it. A rule you cannot trace is a rule you have to take on faith, and nobody gives an AI that.
- You can kill it. When the rule stops being true, you correct it once, in the same place, and it stops being applied.
#What we got wrong
We built the eager version first: a compiler that read everything overnight and produced a tidy summary of the company. It was elegant and it made things worse.
Two findings from killing it, both uncomfortable:
- A cached summary that nobody invalidates is poison. It answers confidently from a world that has moved on. Measured against reading the sources at the time of the question, it was worse, not slower-but-equal.
- Most of what a company writes is not knowledge. The overwhelming majority of the raw material was activity — who did what, when. Compiling it produced volume, not understanding.
What survived was smaller and duller: read at the time of asking, inside the scope, and let corrections accumulate as rules with links back to where they came from.
#The line we hold
A coworker should be able to explain any answer with two things: the source it read, and the correction it applied. If it cannot produce both, we would rather it said it does not know.
That sounds like a limitation. In practice it is the reason people give it the second job.