Knowledge Management: AI Agents
The Knowledge Brain, what it must contain, how retrieval answers "how did we solve this before", and why retirement is the hard part rather than capture.
Markdown. No sign-up, no email.
The problem this function exists to solve is simple to state and expensive to ignore: a company should not lose what it knows when a person leaves.
Almost every organisation solves this by creating a wiki, contributing to it for a month, and then maintaining a searchable archive of things that were once true.
The Knowledge Brain#
One place, twelve kinds of content, each with an owner and a review date.
| Contains | Why it is worth keeping |
|---|---|
| Projects | What we did, for whom, and what actually happened |
| Architecture | Decisions and the options rejected, which is the part that stops re-argument |
| SOPs | How recurring work is done here |
| Policies | The rules, with one authoritative version each |
| Technical documentation | How our systems work, written from the systems |
| Client knowledge | Context, constraints, history. Departure-proof |
| Lessons learned | Including the expensive ones, especially those |
| Proposals | What we offered and what won |
| Templates | The documents we produce repeatedly |
| Research | Findings, including negative results |
| Training | How someone new becomes useful |
| Code knowledge | Why the code is shaped this way, which the code cannot say |
The value concentrates in the "why" rows. What we did is recoverable from records; why we chose it is only in someone's head, and it leaves with them.
Retrieval: how did we solve this before?#
Someone asks in plain language and gets an answer with its sources attached.
Three rules that decide whether this is trusted or abandoned:
1. Every answer cites its documents. An answer without sources cannot be checked, and one wrong unsourced answer destroys trust in the whole system for a year.
2. Superseded material is never retrieved as current. Retrieval that returns a two-year-old approach with the confidence of today's is worse than search, because search at least shows you the date.
3. "I do not know" is a valid answer. A system that always produces something produces plausible-sounding fabrication at exactly the moment the knowledge is missing, which is the moment someone most needs to know it is missing.
The four agents#
Capture agent#
Drafts a decision record or lesson from a meeting, thread or incident, which the owner then corrects and confirms. Nobody is asked to write from a blank page, and nobody signs off reasoning they did not give.
Staleness agent#
Flags documents past their review date, or referencing systems that no longer exist. It proposes; a person retires.
Duplication agent#
Finds two documents claiming authority on one topic and raises a conflict rather than merging them silently. Two current documents on one subject is a defect, and whichever one gets quoted in an incident is the one the other side has not read.
Retrieval agent#
Answers questions from the Brain, with citations, refusing when the answer is not there.
What stays with a person#
- Declaring which document is the authority when two conflict.
- Retiring anything. Effectively irreversible even when technically reversible.
- Deciding what goes into an agent's working context, which shapes every output that agent produces anywhere in the company.
- Confirming a decision's reasoning. An agent can draft it; the person who decided confirms why.
The hard part is retirement, not capture#
Capture is now nearly free. Everyone believes the difficulty is getting knowledge in.
It is not. The difficulty is taking things out, and it is where every knowledge base dies: nothing is ever removed, the proportion that is still true falls year by year, and eventually people stop trusting any of it because they cannot tell which part is current.
In an AI-operated company this stops being an inconvenience and becomes a live hazard, because these documents are what the agents read. A stale page is no longer merely misleading to a colleague who might sense something is off. It is an instruction, followed at scale, without hesitation.