Part 1 - The Write Path Is the Moat

Every AI memory product on the market describes itself on two axes. What the model knows from training, and what it remembers about you. Read their own architecture posts and you will find the same two boxes.

I built MrAgentˣ because the box that matters for real work is missing from the diagram.

Here is the situation that made it obvious. A client's budget was agreed in March. It changed in May. It changed again in July. Every one of those emails is true. Every one of them is in the archive, and every one of them gets retrieved when you ask an AI "what is the budget." Memory that remembers everything you ever said is not memory. It is a transcript with a search box, and a transcript cannot tell you which of three true sentences is still true today.

I tried the usual fixes. Summaries drift. Bigger context windows retrieve all three budgets at once and let the model guess. A popular memory API handles contradiction in a way its own open-source documentation describes plainly: when a new fact conflicts with an old one, both get deleted. The conflict is resolved by forgetting.

So the architectural decision was to stop treating memory as something you read better and start treating it as something you write correctly. The third axis is state: not "the record contains X" but "X is currently true, since this date, replacing Y."

Three things had to be true on the write path for that to hold.

Facts have a lifecycle. Every atom on the Context Map knows when it became true, when it stopped, what replaced it, and when it was last re-confirmed. When the July budget arrives, the May one closes its window and stays on the record as history. Research nobody has re-checked in ninety days is flagged stale instead of served as current. Unknown dates stay honestly unknown, nothing is fabricated to fill a column.

Conflicts stay visible. When new content contradicts the map and the system cannot tell which side wins, both sides stay alive and the conflict itself is returned with your search results, computed on the database side with zero AI tokens. You see the disagreement at the moment of use, which is the only moment it matters.

Every fact carries its receipts. Each atom is stamped once, at creation, with where it came from and whether a pipeline verified it or an AI inferred it. The stamp survives every later update untouched. What you received and what you worked out with your AI never mix, and an AI chat can write its findings into the map without ever blurring that line. The reason this matters is measured: in a study of a mainstream assistant's memory, 96% of stored memories were created without the user asking, and 52% were psychological inferences about the user rather than facts.

Every competitor advantage I have audited is on the read path: faster retrieval, better ranking, bigger windows. All of it sits on top of whatever was written. If what was written is three budgets with no lifecycle, the best retriever in the world hands your AI a coin to flip.

The write path is the moat. It is also the unglamorous part, which is why nobody else built it.

Context Windows Close. AI Forgets Everything. Your Work Should Never Start From Zero.

MrAgentˣ is in private beta. Limited to the first 1,000 until launch.