Every system you own answers a different question: what is true now. That is the wrong question about half the time you ask it.
An LP asks why you passed on a business that has since tripled. You pull up the file. The screen memo is there, and it looks thin, because the company record in your CRM has been updated four times since 2023. The revenue figure on screen is this year's. The management team listed is the one installed after the recap. The sector tag was reclassified in a tidy-up last spring.
You are looking at today's company and trying to explain a decision made about a different one. Every CRM, every document store and every AI tool built on top of them holds one version of the truth, the current one, and quietly discards the rest.
Two Clocks
A fact has two dates: the date it became true in the world, and the date your firm found out. A company's revenue fell in March. Your analyst learned it in June, from a data room, in a document dated May.
Example Record
Northwind Components Ltd
True from March 2026, known to the firm from June 2026. Data room, document dated May.
True from September 2025, known to the firm from November 2025. Company announcement.
True from April 2026, known to the firm from April 2026. CRM, reclassified in a tidy-up.
True from November 2024, known to the firm from January 2025. Press coverage.
True from June 2026, known to the firm from June 2026. Origination note.
On this date the record was complete and current. Every fact had reached the firm. Synthetic data, not a client record.
Where It Stops Being Academic
None of these are hypothetical. Each one is a question a firm has already failed to answer.
To an LP, an IC, a regulator
The business comes back round
When an agent recommends something
Ten passes that turned out wrong
Why Nobody Else Offers This
Bitemporal modelling is old. Banks have done it for decades in transaction systems. It is not new and we did not invent it.
What is new is putting it underneath an AI system that reads your unstructured mess: emails, memos, data rooms, call notes. That is harder, because a document does not announce which dates it is talking about. You have to work that out on ingestion, and get it wrong in a recoverable way.
Most AI tools skipped it because retrieval demos better than history. A search box that returns today's answer looks impressive in a thirty-minute pitch. The cost shows up two years later when somebody asks a question about the past.
One open-source project, Graphiti, does this properly and does it well. It is a library for developers building agents, and if you have an engineering team who want to build the surrounding system themselves, look at it seriously. We are the version that arrives connected to your firm's systems with the workflows already running.
Related: how every answer is traced and who can see which part of the graph.
Pick one from two or three years ago where the outcome is now known. We will reconstruct what your firm actually held about that business at the moment of the decision, and show you what a model would have surfaced. Forty-five minutes, and you keep the output.