What we mean by a memory graph
Retrieval over documents answers what was written. A memory graph answers what was decided, why, by whom, and whether it worked.
Ask a document-retrieval system why a supplier was selected and it will find you the email where someone announced the selection. Ask a memory graph and it should return the decision itself: the candidates considered, the criteria applied, the evidence at the time, the person accountable, and what happened afterwards.
Documents are a lossy record of decisions
The gap matters because documents are a byproduct. They capture the announcement, rarely the reasoning, and almost never the rejected alternatives. Six months later the question that matters is not what we chose but why we ruled out the other two. That information usually existed only in a meeting.
This is why retrieval-augmented approaches plateau in operational settings. They are excellent at surfacing what was written and structurally unable to surface what was decided, because the decision was never a document in the first place.
Modelling a decision
A decision in the graph is a first-class node. It links to the entities it concerns, the evidence it drew on, the policy it ran under, the human who owned it, and above all the outcome once one is known. That last edge is what turns a record into precedent. A decision without an outcome is a claim; with one, it is evidence.
The graph is also where contradiction becomes visible. When two decisions in the same class went opposite ways on similar evidence, that is worth surfacing to a human. It usually means either the policy is underspecified or the context differed in a way nobody wrote down. Both are useful findings.
Forgetting on purpose
A memory that only accumulates becomes unusable and, in a regulated context, a liability. Retention has to be a design decision: what is kept, for how long, at what granularity, and what is deliberately discarded. Personal data has its own clock, set by law rather than usefulness.
There is also a subtler kind of forgetting. Precedent from a company that no longer exists, with a different size, a different market and a superseded policy, should decay in weight rather than sit there arguing for a decision that made sense in 2021. Recency is not truth, but staleness is a real signal.
Why this is the hard part
Connectors are engineering. Execution is engineering. The graph is where the modelling judgment lives, and it is the part that most determines whether the system's reasoning is any good. Get it wrong and you have built a very expensive search index over your own email.