KeyGraph
Memory stays ours.
Models are interchangeable.
Source-bound continuity under real working load. Every fact is tied to its primary document; every high-severity action stops at a person.
Read the architecture →Clever amnesiacs.
Today's assistants reason well and remember nothing. Between sessions the thread is gone. Inside a session, when they don't know, they guess - fluently, confidently, in exactly the same register they use when they do know.
In clinical, professional and other high-stakes operations that is not a rough edge to be smoothed later. It is disqualifying. An answer that sounds right and cannot be traced costs more than no answer at all.
Most AI projects don’t fail because the model is dumb.
They stall on bruises that show up after the first wow demo. Name the bruise first. Architecture second. That is how real buyers talk - and how we write.
“We brief it every morning. By lunch it’s gone.”
Cost: practitioner time, tokens, and a case that never compounds.
“It sounds right, and we can’t prove a claim.”
Cost: quiet risk. No trail for a board, peer, or regulator.
“We added retrieval. It still fills the gaps.”
Cost: false confidence. Similarity is not a cited source.
“The moment it can act, nobody owns the door.”
Cost: blast radius. Chat is wide; write access must be narrow.
“Our memory lives in a chat log we don’t control.”
Cost: lock-in. You can’t swap the model without losing the record.
“The demo worked. Production is a different animal.”
Cost: rework. Fluency is free; being right under load is the product.
Name it in their words → one line on what it costs → what “good” looks like → only then the seat (source-bound memory, human gate, memory ours). Never open a cold conversation with architecture.
Not a smarter autocomplete. A continuity and governance seat for those six pains - always-on, source-bound, verify-don’t-infer, high-severity stops at a person. Why not just RAG →
Five commitments, in plain terms.
Demos reward fluency. Real work rewards being right.
In a demo, plausibility passes. In operations an answer is worth exactly what its source is worth - so the source travels with the answer.
Re-briefing a stateless model is paid twice: once in tokens, once in the practitioner's attention. Memory that persists is the cheaper unit of work.
An agent that can act needs a narrower door than one that can only talk. Write access is scoped, logged and revocable, and severity decides what needs a signature.
Whoever holds the memory and the approval gate holds the system. That should be the institution accountable for the work, not the model provider.
Five parts, one direction of travel.
Cognition ≠ system of record.
The model does the thinking. The graph is the record. Keeping the two separate is what lets either one be replaced without losing the institution's memory.
Most people are only now hearing about RAG. That is the right starting point.
Retrieval-Augmented Generation is the idea most teams meet first: fetch relevant chunks, stuff them into the prompt, answer. It is useful. It is not a continuity seat, not an audit log, and not a place high-severity actions can stop for a person. KeyGraph is built for that second job, and it can sit beside RAG rather than pretend retrieval never happened.
| Dimension | Typical RAG stack | KeyGraph |
|---|---|---|
| Unit of memory | Chunks in a vector index for this prompt | Nodes and edges you can freeze, count, and walk |
| Provenance | “Came from similar text” (often fuzzy) | Every fact tied to a primary source document |
| Across sessions | Start over; retrieve again | Always-on continuity beside the work |
| When the model is unsure | Still fluent; risk on the reader | Verify, don’t infer: no source, no claim |
| High-severity action | Usually outside the pattern | Stops at a person; logged approval |
| Who holds the seat | Often the vendor stack + the prompt | Memory stays ours; models are interchangeable |
RAG can still fetch. KeyGraph is the seat that decides what may be remembered, claimed, and acted on, with residency and custody you can audit. Retrieval is a tool; the graph is the record.
RAG answers this question from files. KeyGraph keeps the institution honest across questions: source-bound, human-gated, memory not rented from the model.
Enterprise knowledge graphs, agent-memory products, and orchestration graphs (workflow engines) solve adjacent jobs. Straight comparisons with those classes come after RAG is clear - same honesty: different load-bearing job, not a claim that one tool wins every bake-off.
Where this actually stands.
Active private R&D, run from Adelaide, South Australia.
Being packaged for high-compliance pilot settings in South Australia.
A public SaaS product. There is no sign-up, no waitlist and no self-serve tier.
Regulated clinical decision support. It does not diagnose, prescribe or treat.
Research Collaborative
Luu Zaibatsu Research Collaborative Pty Ltd
ACN 702 511 037 · ABN 78 702 511 037
Kien Luu, PhD, PharmD