A modern enterprise may hold millions of documents, contracts, invoices, reports, tickets, meeting notes, spreadsheets, source-code repositories, audit logs, customer records, and machine-generated artifacts. Nearly all of it is indexed, embedded, summarized, and searchable within seconds by systems built for exactly that.
And yet, when an organization asks a higher-level question, the weakness shows immediately: what do we currently know? Why do we believe it? Which evidence supports it? Has it been superseded? Which version should an agent or employee act upon?
These are not retrieval questions. They are memory-governance questions, and nothing in the standard stack answers them. Search systems find documents. Databases retrieve records. Knowledge graphs represent entities and relationships. Retrieval-augmented generation assembles context for a prompt. Every one of these is useful and necessary — and none of them, by itself, defines what an organization should remember as operational knowledge, how that memory changes, who or what may rely on it, or how it should be audited over time.
The gap matters more with every quarter that AI systems become operational actors rather than search interfaces. When a model or agent summarizes a report, drafts a contract, reconciles an invoice, plans engineering work, or answers a regulatory question, a plausible answer is not enough. The organization needs a reliable memory substrate that preserves the lineage and status of the knowledge being used — not just its content.
An Addressable Memory Network (AMN) is our answer: a framework for converting heterogeneous source material into governed memory objects that can be persistently referenced, traced to evidence, revised over time, and used by humans, applications, and AI agents alike. It is not a replacement for your database, search engine, knowledge graph, or RAG pipeline. It is a memory layer that sits alongside them.
The core idea is simple to state and hard to build: important knowledge should not remain trapped as an incidental fragment inside a document, a prompt, an embedding, or a generated answer. It should become an addressable object — meaning it has a stable reference within the organization's knowledge environment. Not a phrase buried in a report. Something you can point to, inspect, revise, retire, and audit, the same way a person tracks a decision in their own head, except the system can now do it too: with identity, provenance, lifecycle state, revision history, and governance context attached.
AMN isn't presented as a universal truth system, and it doesn't claim all knowledge can or should be centralized. Its purpose is narrower and more practical: helping a bounded organization preserve, revise, inspect, and act on knowledge without repeatedly reconstructing it from raw artifacts every time someone asks.
The implications reach further than any one organization, though. Once memory becomes governed and addressable inside an institution, the possibility of interoperability follows naturally — independent organizations eventually needing ways to reference each other's governed knowledge across boundaries. AMN starts as a framework for institutional memory. It points toward a longer-term question: whether memory itself can become interoperable, the way documents and data already have.
This framework is what the rest of our research record is built on — the routing thesis we tested and retired, the knowledge graph we measured and found contributed nothing, the confident-miss failure mode we named. Each of those is a story about what happens when you actually test an AMN's components against real data instead of trusting that governed-sounding architecture is the same thing as working architecture.