In part. A shared context file is enough for one person, or for a very small team where everyone may see everything. Beyond that point, three structural limits appear. The file does not know who is reading it, it does not keep itself current, and a change carries neither an author nor an approval. No amount of rewriting fixes any of the three.
The document everyone loads and nobody rereads
One day, somebody at your company took matters into their own hands. Over several evenings they produced the reference document, the one that describes the company, the team, the products and the house rules, and that every AI assistant loads at the start of a session. For a few weeks the answers got visibly better, and everyone said thank you.
Three months on, that document describes an organisation that no longer exists. Two people have moved roles, a product has been withdrawn, a process has been replaced. The file has not moved at all. And every AI in the company recites it with complete confidence, because an AI cannot tell that a text has gone stale. It reads the file, and it believes the file.
Three limits no rewrite can fix
The problem is not the quality of the writing. A static shared file carries three structural limits, and they stay in place even when the document is flawless on the day it is published.
The first limit concerns the reader. The file does not know who is reading it. The intern who arrived on Monday and the managing director receive exactly the same text. That leaves you two options, and both are bad. Either you put everything in, salaries, margins and sensitive projects included, and you expose all of it to every person who opens a session. Or you keep anything confidential out, and the document becomes too thin to help anyone with the questions that actually matter.
The second limit concerns time. The file does not update itself, and its staleness is invisible. Nothing in the text separates the line that is still true from the line that no longer is. A human reader can feel a doubt and go and check elsewhere. An AI feels no doubt, and it answers wrong with confidence, a mechanism we have taken apart in why a connected AI answers wrong.
The third limit concerns the record. A change to the file has no author and no approval. Anyone can edit it, nobody finds out, and nobody confirms that the new version is accurate. The day an AI states something false in front of a customer, you can neither trace where the error came from nor say who introduced it.
What replaces it, and why the result changes
Those three limits sketch the requirement in negative. A useful company context knows who is reading it: each person, and each AI working on their behalf, receives only what their responsibility entitles them to see. It is maintained: every piece of information carries its date and its owner, and what has gone stale can be spotted instead of passing itself off as true. And every change goes through human approval, with an author and an approver on record.
This is precisely what Kastel does. Kastel is not an AI, and it does not replace the ones you already use. It is your organisation’s context layer, the structured memory that the AIs you connect come to read, each one limited to what its user is allowed to see. You decide what each AI is allowed to see, and the difference between simply sharing and actually governing is the heart of our approach to context governance. The product page walks through the mechanisms one by one.
When a single file is still the right tool
The honest answer to the question in the title really is “in part”, so let us say for whom the first part holds. If you work alone, a single context file is the right tool. A freelancer or a solo founder has no confidentiality gradient: nobody else reads the file, and keeping it current depends only on their own discipline. A team of two or three people who tell each other everything can live the same way.
The threshold is crossed with the first document that not everyone may see. A salary grid, a per-client margin, a hiring plan or a possible sale of the business will do, and one single case is enough to make the shared file unworkable. From that point on you either expose or you impoverish, and no rewrite gets you out of that choice. What happens once that threshold is behind you, and what moving on looks like, has a page of its own: when a context file stops being enough. The neighbouring question, what an assistant remembers about you personally, deserves its own treatment, and we have written it up in company context and personal memory.
Check it without taking our word
You can test everything above on your own machine. The core of Kastel is free, installs on your own infrastructure with no size limit, and connects to your own AI keys. Build your context, then deliberately let one fact expire: change an owner, retire a product from the catalogue. Ask the AI and watch what it cites: the version in force, dated, rather than the recollection of a file nobody has re-read. Then export the whole thing: all of it belongs to you, and all of it comes out.
Kastel is not exempt from its own rule: full export at any time, free self-hosting for life. You can check every one of these claims yourself.