Ask an enterprise AI vendor what "memory" means and you will usually get a description of retrieval: embed the documents, index the chunks, pull the nearest fragments into the prompt. It works, in the way a filing cabinet with a fast intern works. It is also a category error about what organisations actually need to remember.
Consider what a document is: the surviving artefact of a decision process, written by the winner, after the fact, with the disagreements sanded off. The pricing page shows the price. It does not show the three models that were argued over, the churn data that settled it, the objection from finance that turned out to be right, or who actually made the call and on what authority.
Documents are what a company decided to say. Trajectories are how it learned to think.
When a new hire, or a new agent, faces a related problem, the artefact is nearly useless. The value is in the reasoning that produced it: what was tried, what failed, what constraint bound, what the world said afterwards. That is the judgement a company pays for, and it is exactly what artefact-centric systems throw away.
The trajectory as the native unit
A trajectory is a governed record of a piece of work as it actually unfolded: the question that opened it, the evidence gathered, the alternatives considered, the decision and its authority, the action taken, the outcome observed, and the revision when reality answered back. Seven beats, linked with provenance, on a record that never rewrites its own history.
Two properties matter enormously and are easy to miss.
First, trajectories are captured during normal work, not documented after it. Knowledge management failed for thirty years because it asked busy people to do homework. A harness that rides alongside people and agents inside the tools they already use gets the record as a by-product: nobody writes anything up.
Second, trajectories are governed. Every element carries who may see it, where it came from, when it expires, and how it can be revoked, with receipts for every access. That is the difference between an asset an enterprise will actually deploy and a liability its security team will kill on sight. Memory without governance is a leak; governance without memory is a filing cabinet.
Why retrieval cannot get there from here
Could you bolt this onto RAG? Not really, and the reason is structural. Retrieval treats knowledge as text to be found. Trajectories treat it as causally linked judgement to be inherited. A vector index cannot tell you which alternative lost, or that a decision was superseded last Tuesday, or that the person who wrote the memo was overruled: the text never contained any of it. You cannot retrieve what was never stored.
The payoff for storing the right unit is not subtle. The next mind facing a related problem loads the thread, not the corpus: cheaper on tokens, faster to useful, and immune to the confident reconstruction of context that never happened. And because trajectories live outside any single model, the judgement survives model churn: swap today's frontier model for next year's and the company's memory comes along intact.