Institutional knowledge capture: a system that does it daily
Capture is almost always organised as an event, and knowledge does not arrive as an event. The version that survives a busy quarter is the one nobody has to decide to run.
Most institutional knowledge capture is a project with a start and an end. An offsite where everyone writes up their process. A documentation push before the quarter closes. Forty-five minutes with a departing colleague and a form. Each is a genuine attempt, and each shares one defect: knowledge accumulates every working day and the capture happens twice a year, so what you end up holding is a snapshot of a company that has already moved on, assembled by people guessing at what will matter later.
The alternative is not a better event. It is capture as a byproduct of work that was happening anyway. Mail gets read. Calls happen. Documents arrive attached to threads. If something reads that stream and writes the record out of it, capture stops being a task competing with the work, which is the only property that has ever let it survive contact with a bad month.
How to capture institutional knowledge
The mechanism is smaller than the phrase suggests, and there are three moves in it. Something arrives and is read, then filed against the program, deal or agreement it belongs to. The record for that thing is updated rather than appended to: the items the message resolves get closed, the items it creates get opened, the dates it moved get moved. Then the reversible next step, usually a reply, is drafted and left for you to approve. Nothing goes out on its own.
Capture is the middle move, and it happens because the first one did. Nobody sits down to write a page. The cadence is a scheduled pass rather than an instant, hourly at best, and it needs a machine that is awake for the pass to run on. For this problem that is more than enough, because the failure being fixed is measured in months rather than minutes.
What makes this different from filing is that the unit being captured is state rather than content. A document management system holds the study report. Capture in this sense holds the fact that the report is now due three weeks later than the contract says, that the reason is a delayed shipment, and that two downstream items depend on it. The general form of that argument, at the level of the whole record rather than the individual item, is AI institutional memory.
What a week of capture actually collects
| What arrived | What is worth keeping | Where it sits if nobody captures it |
|---|---|---|
| A contract research organization moves a report date by three weeks | The new date, and the downstream items that move with it | One inbox, until somebody next opens the timeline |
| A partner says on a call that the data transfer goes out on the 23rd | A commitment you are owed, with a date on it | The memory of whoever was on the call |
| An executed agreement with a notice period in clause 8 | An obligation whose clock starts on an event, not a date | A signed PDF nobody has opened since signature |
| Two vendor quotes and the decision between them | Why the more expensive one was chosen | Gone the moment the choice is made |
Nothing in the middle column is a document. Every item is a small piece of state with a date or a reason attached, which is why capture organised around writing pages never reaches it. Nobody writes a page about a date moving, and nobody would read it if they did. This is also the honest limit of the approach: it captures the stream that reaches one seat, one mail identity, one owner, so it holds what passed through your correspondence and calendar rather than everything the company knows. The company-wide version of that is several personal records rather than one shared brain, and the gap between those two things is the subject of a second brain for companies.
Capture knowledge before employees leave
The search that brings most people here starts about two weeks before somebody's last day. By then the window has closed. What can be done in a notice period is triage, and triage is worth doing, but you cannot reconstruct the reasoning behind two years of decisions in a conference room, from memory, with a person who has already mentally left. The capture that rescues you happened in the eighteen months before the resignation letter or it did not happen at all.
Which turns the useful question around. It is not what to ask on the way out, it is what was being written down all along. If the record already holds the commitments, the reasons and the dates, a departure becomes a handover of a file rather than an interview, and what that handover consists of is worked through in institutional knowledge transfer.
Why external correspondence makes this easier, not harder
A drug development company generates surprisingly little of its own state. The tox study is at one contract research organization, the analytical method transfer is at another, the manufacturing slot is at a third, the partner runs a steering committee that meets quarterly, and the regulator has a clock. Almost everything that decides whether the filing date holds arrives from outside, from people who will never touch your tracker. That is what defeats capture built on internal discipline, because there is no internal moment where the fact gets typed in.
It is also why byproduct capture works unusually well here. Facts that arrive from other organizations arrive already written, as email, as attachments, as minutes, into one account. The stream that has to be read is your correspondence, and that is exactly where the state lives. The version of this argument written for the regulated half of the industry, and the line between the two, is knowledge management in pharma; the discipline-level framing is AI knowledge management for a drug development team.
Capture nobody has to schedule
A quote chosen over the cheaper one, a transfer the vendor put on the 23rd, a notice period sitting in clause 8: each gets written into the record off the correspondence that carried it, with no page authored by hand. The demo is a record built exactly that way, months into running.
See the demo