They are objects of a different nature. An assistant’s personal memory is what one tool has learnt from one person over its conversations. Company context is the shared reference, built from the sources that carry authority, with owners, read scopes and a history of changes. The first never grows into the second, and mixing them lets confidential facts escape all governance.
Nobody can say any more which AI knows what
A managing partner’s AI assistant picks up a great deal over its conversations. It has learnt how she writes, her turns of phrase, the formats she prefers. One evening, while she was preparing a meeting, it also retained the acquisition project under discussion. Her fellow partner’s assistant knows nothing about it.
Six months later, nobody in the organisation can say which AI “knows” what. The acquisition project lives somewhere inside a tool’s memory, attached to a personal account, and nobody ever decided to put it there. The day it needs to be forgotten, nobody knows how to make that happen, or even how to verify that the forgetting actually took place.
Two objects of a different nature
An assistant’s personal memory is reactive. It retains whatever happens to pass through the conversation, at the moment it passes, with no plan and no validated selection. It is tied to one tool and to one person. On the company side it has no owner: nobody answers for it, no read scope applies to it, and its forgetting procedure, where one exists, cannot be verified.
Company context is built the other way round, from the sources that carry authority: the organisation as it actually is, the processes in force, the decisions that were taken. It has owners who answer for it. Every person reads it through a scope that follows from their responsibility, and every change leaves a history, with its author and its approver.
That is why growth never turns the first into the second. No accumulation of individual recollections produces an approved shared reference, because the first object lacks what defines the second: construction from sources, ownership and approval. A stack of private notebooks, however tall, never becomes a register the organisation stands behind.
What happens when the two get mixed
Go back to the acquisition project from the opening. Once it has entered an assistant’s personal memory, that confidential fact escapes all governance. No read scope applies to it, because that memory knows nothing of the organisation’s responsibilities. No record says who can reach it, nor in which answers it will resurface. And when it becomes false, because the deal has been dropped or closed, nobody can correct it, since nobody knows it is there.
Mixing in the other direction does damage too. If the shared reference is assembled from whatever the assistants happened to retain in conversation, it inherits their blind spots and their unverifiable claims, and governing that context becomes impossible, because nobody can say where each statement came from or who approved it.
The practical rule, and where Kastel sits
The boundary fits in one sentence. Personal memory keeps one person’s working preferences: how they write, the formats they favour, their shortcuts. Company context keeps the organisation’s facts: who does what, what was decided, what is in force. A fact that concerns the organisation should never live only inside a tool’s memory, and an individual preference has no place in the shared reference. The moments where that boundary gets paid for, a new joiner, a resignation, a change of tool, are gathered in what to leave in an assistant’s memory and what to take out.
Kastel takes on the second half of that boundary. Kastel is not an AI, and it does not manage your assistants’ memories. It is your organisation’s context layer: the shared reference, structured from your sources, with a read scope per person and a history of changes, which the AIs you connect come to read. The product page describes these mechanisms. If your starting point is a big shared document instead, we have examined its structural limits, and the question of where your company memory should live has a page of its own.
What you can verify for yourself
None of the claims above asks you to take our word. The core of Kastel is free and installs on your own infrastructure, with no limit on size or duration. Connect your own AI keys, build a first context from your sources, give two people two different read scopes, and watch what each assistant can read. And if you ever decide to leave, you export your entire context and take it with you: it belongs to you.