Part 5 - The Before and After

I ran the full pipeline last week on 135 emails, documents, and saved AI discussions.

One honest disclosure first, because this series has been hard on systems that blur what happened with what was staged. I cannot show you a client's inbox, mine included, without breaking the confidences inside it. So I built one: a corpus of 135 emails, documents, and saved AI discussions modelling a working life. An acquisition closing. An invoice dispute. A rebrand kicking off. Medical follow-ups. Vendor threads. Working sessions with an AI about the decisions in them. The names are invented. Every number that follows is the real system's real output on that corpus, and none of it was touched after the run.

Before

The corpus looked like any working week looks. Ninety-eight emails, thirty-five documents, and the AI discussions worked out alongside them, each one smaller than an email and easier to lose. Commitments buried in paragraph three of replies. Decisions scattered across threads with different subject lines. Risks mentioned in passing, acknowledged with a "good point, let's watch that," and never seen again.

None of it visible on the surface. That is the point. The surface always looks fine.

The processing

I forwarded the corpus and waited.

The ingest took thirty-four seconds. The extraction queue then worked through all 135 documents on its own, unattended, while I did something else: about an hour of processing in practice. No tagging on my side. No folders. No decisions at the door.

After

The map that came back was that working life, structured.

654 atoms extracted: 6 commitments, 56 decisions, 54 risks, 120 tasks, 350 facts. Each with its source document and its date. And the atoms born from the AI discussions carry a different stamp than the atoms born from the emails: what was received and what was worked out with an AI, on separate tracks, visible a year from now.

47 topics assembled themselves. The acquisition pulled five documents into one thread of record. The invoice dispute pulled four, from first invoice to current standoff, without anyone filing anything.

Of the six commitments, five are open. One of them reads, in full: "November 1 customer communication date is firm." A hard date, agreed in writing, attached to no calendar, owned by nobody. In a real inbox that sentence is a small bomb with a two-month fuse. Here it sits in a list, dated and sourced, where someone will see it before November does.

Fifty-two of the fifty-four risks had been mentioned and never tracked. They are now a list with sources attached.

And one finding I want to report precisely because it is unglamorous: the run found zero contradictions. The system checked; the corpus happened to contain none. A memory that reports "no conflicts found" is making a verifiable claim. A memory that never checks is making no claim at all. I would rather show you the boring true number than a dramatic invented one.

The numbers

135 documents processed, saved AI discussions among them. 654 atoms extracted. 47 topics clustered automatically. Ingest in under a minute; extraction about an hour, unattended.

Time to build the same picture by hand: days. Realistically, never. Several of these connections only exist when the whole map is assembled in one place, and no human assembles their whole map.

What this means

Notice everything the system did not do. It wrote no emails. It scheduled nothing. It generated no content.

It preserved operational state. That is all. And it turns out that was the missing thing, the layer everything else was supposed to be sitting on.

A working life became a map. The next run will be on mine, and the run after that on yours.

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.