'Governed memory' is about to mean five different things. Only one of them is ours.
The phrase governed memory is about to get borrowed by an infrastructure primitive, a data-governance catalog, a customer-memory system, and an academic architecture, each a legitimate use of the same two words for a different problem. Here, on SAGE's own structural terms, it means one specific thing: provenance on every fact, permission-aware recall, contradictions resolved when provable and escalated to a binding human ruling when they're not, and a memory that never leaves the customer's own infrastructure.
Governed memory was coined to mark one exact claim: an operational memory sitting under real enterprise AI, where provenance, permission, and contradiction are structural properties of the system, not promises in a slide deck. That's still what we mean by it. But the phrase is two ordinary, useful words, and useful words get borrowed. Expect governed memory attached to at least four other ideas over the next year. None of those uses are dishonest. None of them are this.
A word under load
An infrastructure primitive will call itself governed because it bolted an access-control layer onto a session or context store. Permissions are the governance most infrastructure teams mean by the word, and that's fair enough. A data-catalog layer will call itself governed memory because it maps lineage and ownership across a data estate, a real map, but of data sitting at rest, not something an AI queries live mid-answer. A customer-memory system will call itself governed because it logs consent alongside what it remembers about a person's preferences: memory of the customer, not memory of the business the AI is meant to be reasoning about. And an academic architecture will describe a memory module's governed behaviour as a contribution to a field, a real design that nobody has yet pointed a live AI tool at inside a real company's infrastructure. All four uses are honest. None is what SAGE does.
What governed means here
Here, governed isn't a permissions checkbox sitting beside the memory. It's a property of every fact inside it: each one traceable to the document, version, and ingestion that produced it, so any answer built from it traces back to a source a person can open and check. It holds the same standard when two facts about the same thing disagree, not by quietly keeping whichever is newest, but by checking whether lineage, supersession, or an organisation's own authority ranking settles which one is current, and logging that decision either way. That's what governance costs when it's structural: a receipt on every fact, a ruling on every conflict.
What memory means here
Memory, here, isn't a catalog sitting in a repository for someone to search when they remember to. It's the layer every connected AI tool queries live, permission-checked at the moment of the question rather than the moment a document was filed, so a fact nobody cleared this requester to see simply never surfaces, whichever tool is asking. And it's one memory of the business, not one per tool. Decide something once and the whole fleet inherits it, not the instant it's decided, but the next time any connected tool asks.
The two-tier test
Most definitions of governed treat a contradiction as an edge case to flag and move past. Here, it's the test of whether the governing is real. Where lineage, supersession, or an authority ranking settle a conflict outright, SAGE resolves it without a person and logs the resolution: arithmetic, not judgment. Where none of those settle it, two facilities, say, each updating the same procedure after the same regulation changed, neither referencing the other's revision, SAGE escalates it to a person, whose ruling is written back as its own record and becomes what every connected tool receives from then on. Disagreements don't vanish under this test. A governed memory that hid its disagreements would be worse than one with none: it would be hiding exactly what an auditor most needs to see.
Where it has to live
A memory holding everything a regulated organisation knows isn't something to hand to someone else's cloud, however governed the dashboard looks. SAGE deploys entirely inside the customer's own infrastructure, the same walls the AI tools querying it already sit behind, because sovereignty isn't a feature bolted onto governance afterwards. A memory that leaves the building has already failed the provenance and permission tests, whatever else it gets right.
None of the other four uses of governed memory are wrong. Each is solving a real, adjacent problem with the only two words English offers for it. But provenance on every fact, permission-aware recall, a working test for what gets resolved and what gets escalated, and a memory that never leaves the customer's infrastructure: that combination isn't a marketing flourish sitting on a database. It's the actual shape of the problem a regulated organisation has, and it's the only one of the five meanings SAGE was built to satisfy.
Loriq builds SAGE, the governed memory engine. Talk to us.