Memory
The organisation's shared knowledge and each project's memory, and how labels decide who may read what.
Agents work better when they know what the organisation already knows: its handbooks, standards and decisions, and what each project has learned. AIOE holds that knowledge as the index of record, in two places, and answers every read only as far as the reader may see.
| Memory | What it holds | Where |
|---|---|---|
| The organisation's | Handbooks, standards and decisions filed for everyone, and what projects share into it. | Memory |
| A project's | What the project's workbenches extract from its repositories, and the notes, lessons and decisions agents write back. | The project's Memory tab |
The organisation's memory
The Memory view shows how many items are held and of what kinds, where they came from (filed directly, or from a project), and which projects mirror into it. Anyone signed in can search it, and sees only what they may read. Administrators write to it; in self-hosted AIOE, so do roles granted memory:write under Settings, Roles.
Share a project's memory with the organisation
A project's owners decide what it shares, on the project's Memory tab under Share with the organisation:
- Keep sharing as it changes: Off, Notes and decisions, or Everything. The organisation's copy is kept in step on every change.
- Share now: copies what you can read now, once.
A shared copy never has a lower label than the project's, and it names the project it came from.
Ask Microsoft 365 too (self-hosted)
An administrator can switch on Microsoft 365 grounding under Settings. From then on, a question put to the memory is also put to Microsoft 365 Copilot through Work IQ, as the person asking, and the cited answer appears under From Microsoft 365 beside the organisation's own, with the sensitivity label Microsoft put on it. Nothing from Microsoft 365 is copied into the memory, Microsoft's permissions and labels apply to what comes back, and Microsoft bills each question in Copilot Credits. A follow-up question continues the same conversation.
It needs three things, which the panel spells out: Work IQ enabled for the organisation with a Copilot Credits spending policy; the delegated permission WorkIQAgent.Ask on the API's app registration, with admin consent; and the platform's own Entra secret, so it can ask on a person's behalf. Only people can ask, never machines: an agent that wants Microsoft 365 context uses the Microsoft 365 (Work IQ) tool server instead.
Who may read what (self-hosted)
In self-hosted AIOE, every memory item carries a label, and every read is filtered by it before anything is ranked or summarised. Hidden items are left out, and so are links that would lead to them. Compartments and clearances are not offered in Mojo Up AI Cloud.
Sensitivity. Each item is public, internal, confidential or restricted. An item with no label is confidential on an enterprise project, internal on a personal one.
Clearances. Everyone reads internal and below. Under Memory (or Settings), Clearances, raise or lower a person or a group; a level set for a person wins over their groups' levels.
Compartments. An item labelled with a compartment (for example customer:acme) is read only by people who hold it. Under Compartments, create one and grant it to people, groups, or every member of a project.
What a person sees. On the Memory view, find a person to see the clearance and compartments they read with.
Labelling rules. A project's owners set rules on its Memory tab, matching repository paths or sources to a sensitivity and compartments. The project's machines label what they send with these rules, and anything no rule labels is treated as confidential. Enterprise projects always block content the classifier cannot label.
Relabelling. A person can change an item's label, with a reason that goes into the audit trail. An agent cannot.
Grant review. A project's Memory tab lists compartments granted and never read over a period, so you can take back what nobody uses.
What is recorded
Every read of memory is recorded in the audit trail by label and count, never by content. So is every refusal, and every relabelling.